Drop Your Thoughts Below And Join Me Live Wed 5pm PST. 🟡 Become your Own Record Label: www.topmusicattorney.com/becomeyourownrecordlabel 🟢 Free Split Sheet Template: www.topmusicattorney.com/splitsheet 📄 Music Contracts Templates: www.topmusicattorney.com/music-contracts 📺 Get Your Music Into Tv/Film/Games: www.topmusicattorney.com/sync 💻 #1 Resource For Artists And Producers: www.topmusicattorney.com ⚖ Hire An Attorney: www.delgadoentertainmentlaw.com 📰 Get My Newsletter: www.topmusicattorney.com/newsletter/ 🎧 Listen To The TMA Podcast On All Streaming: www.topmusicattorney.com/podcast
You are a bit confused on lots of points about machine learning. Are we going to relitigate the idiocy of mp3 players having to make copies in memory of mp3 files in order to play them and that that violates copyright? Trash all the iPhones! When tech illiterate lawyers run up against sophisticated technology, the results are often hilarious.
Thank you for this wonderful channel. I am a song writer that has recently discovered SUNO. I have never found a suitable singer or band to do justice to my songs, but with SUNO I have made excellent DEMOS of my work. I sent some of those to a singer who I seen on "The Voice" program. He sent a message back saying that he worked for a company that detects "AI FRAUD music". His reply was brief, saying that he was reporting me to the "authorities" and that I would be charged. I find this laughable as I personally wrote all the lyrics and music, and simply made a demo, but never released that under any false pretences. This makes me wonder if SUNO should be banned or simply have restrictions placed on it to say, - Yes, you can use AI to make demos or ideas or personal music etc, but not that it can be released on public avenues for generating revenue. Then the problem would be that a band uses SUNO to create a 'copy-style' sound in the arrangement of instruments and the way the singer uses his voice, and then simply copies this with a few alterations which then make it an original. What do you think about this? By the way, if I can find a suitable singer, this is what I would do, and then I would definately require your services to create my own record label, and represent me when the big record labels come after me.
Music can sound similar. Same chords used, similar melodies. This is about the big labels wanting to stop people using new tools to express themselves. Suno isn't copying any more than any human that has listened to music.
Imagine you're a smaller independent artist. You've had a couple big hits, but the rest haven't been that great, but you truck on, keep making music and hopefully will strike gold again. Then... Suno trains itself on your stuff, in particular your most successful tracks... now with a text prompt, anyone can generate very similar tracks to your big hits. They're similar, but different enough to be classed as original mixes. They get released and do well. How the fuck does that make you feel as the original artist now?!? You're painting this like some big evil major labels versus the young upstarter Suno. But you're looking at this all wrong. The people that really get hurt in the end, is us artists.
@@1wibble230 Big studios dont care about artist's rights, they care about money and being without rivals. If they are going to shut down AI music creation, they are going to make their version.
@@1wibble230 I do think that if Suno loses the case, it means that the devil wins. Music is for everyone, this kind of evolution of music making is inevitable.
When the copyright office talks about "derivative works", it's not talking about a producer's tag, or a riff, or drum beat, it's talking about somebody taking a copyright protected work and making any kind of baby image of it, which means the same subject, structural, hook, chorus, or any way the spirit, or meat of the songs, which a lot of Artists have been doing for a long time
"What the AIs are actually trained on is the structure of the information, not the content itself. The AI models learn patterns, rules, and frameworks that allow them to replicate similar content based on that structure. So, the similarities lie more in the underlying structure than in the content itself. I hope this perspective helps put the situation into better context for those following the discussion."
Not quite. The AI *has* to be trained on the content to be trained at all. Part of the training is comparing the model's output to the source, finding the delta and back pedaling to try again with improved knowledge and classification. It's basically advanced pathfinding, with the goal being bohemianRapsody.mp3. :) Through the legal process we now know Suno trained it on data mined music without permission. Suno's defense right now seems to be they don't think they need permission.
@@GoranBackmanMusic When you learn to play a piano you have to be trained as well as a human on all the masterpieces. Then when you can already play and create your own music your beain uses all what you have learnt creating something new. It works quite similiar with AI. In addition, if AI will be trained only on bad music or old music (without copyrights) that means this tool will be no more usefull and will be only a toy.
@@GoranBackmanMusic To be fair, discovery has not happened and your limited explanation on training may be flawed. Any training data has to be converted into tokens, the tokens are going to be portions of the data that are labeled. For instance what notes are played, what instruments are played what lyrics are being sung where are you in relation to the rest of the song(this is a reduced version of possible data points used for training) Now when you train this, you train to the next chunk of tokens i.e. the next portion in the song lyrics to gage your training loss as to what the model thinks will happen vs what did happen. But remember it is not on the song itself, it is on what the next logical outcome would be i.e. next note, next bar, next vocal. Now if you want robustness in your outputs you will generalize and randomize the data that is fed into the machine so the output will be better at generalizing the output as well. Its going to be hard for the record labels upon receiving the training data used to be able to point out any given song portion and then say that adding in additional labeling along with the song data is not transformative in even refining the training data for the model. There is also the NLP portion that is included in the training data that will throw everything off as well.
@toluduggan137 I cant agree with that Because they wouldn't have used it If they didn't need it and the fact that they used it represents infringement in some form..
In the training process a copy of the data is not made; instead a latent space understanding of it is built. You cannot extract the exact source song out even if you tried. You could intentionally attempt to forge a counterfeit but that would be a crime you are committing; not Suno. Imagine: You could use FL Studio or Sound Forge to counterfeit a song but neither ImageLine nor Magix are liable for the crimes - you are. Yes, their software CAN actually produce identical copies and infinite variations of all the copyrighted music if you 'prompt it' correctly but that point is moot. Yes YOU would actually be breaking copyright laws if you did but that point is also moot. This is NOT what is happening with Suno. If you could prove that for instance, Suno produces derivative work, you're still not going to win. Once you prove the chain of derivative works has no end and that each artist actually stands on the shoulders of other artists you'll realize that to win this argument is to ensure that creators will never profit from music ever again. Instead once all the dues are paid to all predecessors the creator will get 0.001 cents per million plays and the only people making money are the litigating attorneys preying on creators to sue other creators into oblivion for 2 bars of a song that 'sorta' resembles a riff they never even created in the first place ^^.
Where that falls apart is that you are generating audio, not just a composition. If it were like FL or any other DAW, you could bring in your own copyrighted audio but that's completely on you. If you remake it, you are intentionally remaking that audio. With Udio and Suno, the compositions are largely original, but the audio is more like stem separation where you can copy a specific artist's voice, producer tags etc. That's the real problem. Even if I were to put in copyrighted lyrics using the same tags as a band it should sound like another band (that probably doesn't exist) in that genre playing the similar song that's been created. I shouldn't hear specific producer tags, riffs from a specific recording of a song, or anything that sounds like the actual training source. That is the major difference between AI and humans. We play / program instruments we paid for or own. The instruments AI is playing to perform the compositions is recreated stems/samples from real audio combined in a way that makes it harder to identify... which is kind of like a remix? Either way, the sources used to train data should have been sourced with consent/ potentially payment if it's not a nonprofit.
@@kerryL Humans accidentally recreate songs all the time; there's literally a bunch of YT shorts that keep popping regarding that scenario (i.e. "Deadmau5 accidentally recreates Sandstorm"). In fact there are also instances of human invention that happened simultaneously on opposite sides of the globe and conflict over who actually invented said mechanism. Thus it is entirely possible - and expected - that there could be some accidental overlap even without intent to do so in AI performed music. As far as the other points: /// You don't have to own an instrument to play it. People borrow instruments all the time and whether or not you own the instrument is not a qualifying factor in whether or not you own the music you produce with it or whether you are allowed to capitalize on said music. /// Humans often strive to exactly mimic the vocal qualities of artists. For instance are you not familiar with Elvis Impersonators? /// Consent... well I think you lose that argument immediately because people are allowed to record anything they can observe in a public place. The internet is a public place thus I can record anything that I encounter (granted as long as it's not in a private area). That does not give me the right to re-broadcast it and claim it as my own - but it does give me the right to learn from it, to run machine code against it to train a model, to utilize that model or to create new music using that model. /// In public you have no expectation of privacy. /// If you place an item in public view your permission for other people to view (i.e. process) said item is implicit and assumed. You have already consented by placing it in the public spotlight. /// This really isn't even about AI artists claiming to have written popular artists music - what this is about are popular artists laying claim to music THEY NEVER WROTE and denying new artists using the latest technology from gaining equal footing in an already unfair market. Traditional artists are attempting to maintain a monopoly on music and that's explicitly forbidden. /// The more I think about that the more I think this is a matter of racketeering, segregation and discrimination - not by Suno - but by their opponents.
@@BeyondPC A lot of your points don't really contradict mine. They are just adjacent. 1 - Recreating songs... even if Deadmaus accidentally recreated derude, the performance / production was still different from the original. I would expect there to be a lot of overlap/recreation, but it's generally social context that makes that more or less significant. Then again, I never said composition was a problem here. Just the audio. 2 - Ownership/consent - Yes, but if you are borrowing an instrument that's like implied *consent.* If you steal someone's 5k instrument to record a hit, they can't really do anything to you legally if you return it promptly without damages, but we're venturing into a morally gray area. If you are going to use someone else's voice, their guitar recordings, and everything else in your finished works and profit off of that, they should be compensated. If it's all free and experimental, there's not really any serious harm there, so long as there is no impersonation. Impersonating someone in an unironic was is also considered morally shameful. It should either be done as a tribute, incorporated into your style, or as a joke. But if I'm using _my_ voice to sound like Elvis, I'm still using my own voice anyways. No harm as long as I'm not telling people I am him. 3 - Public Domain - the normal rules of public domain can't really apply to new technologies. When we first dealt with sampling we had to evaluate it's transformative you and establish new rules around it... only a certain amount of seconds of another work can be used before royalties are to be paid. Also the transformative nature described how samples use a clip of songs or instruments as a new instrument for the human to recombine/remix rather than the performance. You are no longer listening to the singer from the sample or the drummer from the sample. You are now listening to the likeness of the producer creating a soundscape or the rapper/singer adding words over that. That is transformative. Because the producer is not pretending to have created what the singer created their likenesses are separate. The AI tools aren't working in this way. It's been "Oh look this sounds like [insert artist name here] is singing, or [name] played background guitar, or [produced this]..." and it's because they probably did. The composition is original but the performance is a recombined hodge-podge of others. Not og. The composition is original but the audio is not. With a new technology, that requires new rules to compensate others properly. Just like sampling needed. 4- Consent isn't about the recording of something as much as it is about profiting off of it musically. You're mixing musical recording with general recording. You can record things out and about, but if you record a copyrighted song in totality in a coffee shop and post that, you're probably getting a takedown or getting sued. Even though that was in public. Way too oversimplified. Anyhow, I'm not a huge fan of intellectual rights, but I do feel that people should be compensated for their work and likeness. Again, I don't want the labels to win. They don't compensate artists properly anyways. I just want artists to be compensated when their work is used to train neural network models. And especially *audio* models. I don't think what Suno and Udio are doing is inherently bad, but I do think it's not being executed properly if you consider the fact that they profit off of using others voices in this way. If it were nonprofit? A little sketchy but cool. Needs more rules to not be misused. If artists were paid for contributing to training data? Not sketchy. Super cool. If artists were paid for derivative works made using their datasets? Doesn't seem possible really... But it would be good that they can make a consistent living off of extra royalties in fractions of a cent I guess. Probably not worth keeping track of and managing. A little overboard. Otherwise both AI Music Generators and Label Companies are crooks here.
The interesting part is that the studios are reprocessing vocals with pitch correction and autotune actually turning artists into AI voices. Once the voice has been processed this way it is no longer the real natural voice of the artist. They are doing this in studio recordings and in live performances. Trying to sue Suno is actually basically saying that they have the copyright to every sound of every instrument, chords and chord progressions all the notes and all words. It is like suing Wikipedia due to its historical information. How can they positively define what part of Suno belongs to which record company or artist? They can not quantify anything to calculate a financial loss. It's like suing someone for using the same chords in their song as say for example in The House of the Rising Sun. If this were possible no future songs could be made not by Suno Udio or anyone from now on. Music is a language the language itself can not be copywritten and therefore can not be accused of infringement either. The record companies have done this to numerous things in the past, For example, they say karaoke is not licensed for broadcasting or public performance. If that were so then Karaoke itself the whole concept would be unlawful as most of the time that is what you do you publically go and sing. If you can not go to a Bar and publically sing then karaoke would not exist.
Agreed... "How can they positively define what part of Suno belongs to which record company or artist?" - since it appears to be a class action suit brought about by the conglomeration of all the big record companies, they could say they own 99% of all songs ever recorded. "They can not quantify anything to calculate a financial loss." - they can't which is why I suspect they are going after SUNO and not the creators (at this point). They will argue that the music was used to train the AI and that the use of that music was done without proper permission. They will lose that argument IMO. Will they sue the record player company (or iTunes) because Bruno Mars or the Weeknd played records to learn an R&B style? This will end up either with a settlement for licensing rights and/or the platform will be bought out by the record labels. Reminds me a bit of when George Harrison was sued for My Sweet Lord sounding like the Chiffons "He's So Fine" and ordered to pay $1.5 million to Bright Tunes... but then bought the publisher nullifying the entire lawsuit. I suspect the record labels will pay millions... maybe a billion... and acquire t he technology to produce their own new music without any new artists. Talk about anti-trust.
@@GeoTactics Today I made a karaoke song of a copyright song which the words are word for word out of Psalms 23. Now surely the bible can sue them for that lol How can they have established copyrights on the biblical verse they have used.
Just noticed your surname Delgado I am product of family Delgado Borges on my father's side so hello cousin lol. By the way my channel got demonetized 3 days ago accusing me of reusing content long story short they are referring to something I did that is certainly not reused content as it is unique on youtube so as I am just totally fed up with them I have placed all my work in private because even though they demonetized me they were still placing ads on my stuff.
Music companies already rip off most of their creators! They should not be allowed any more control of how music is made! If a song is made by AI it should still be just as relevant for use as any song created in any other way! Even most modern music can be traced back to parts that were used in classical music from long ago! I would bet that if you looked closely , they have overstepped their bounds and the music for many songs is historically public domain! And though there are court cases ruling contrary with songs like My Sweet Lord etc use patterns that are well established as musical fundamentals! If that were not so we would not even have what is called genre! We can classify music because of the similarities and methods used! No one group should have the rights to them! They should be for everyone!
The thing is that chord progressions and rhythms alone cannot be copyrighted. Melodies and lyrics can. So, I can analyze, for example, a Taylor Swift song...look at the stylistic elements--the chords, rhythms, song structures, etc. and replicate those elements to create new song. That is kind of what the bands did during the British Invasion with the Beatles. You had Herman's Hermits, Gerry and the Pacemakers, Chad & Jeremy, Freddie and the Dreamers, the Byrds, the Hollies, and later, Badfinger (who were on Apple Music and were produced by the Beatles and played some tunes co-written by Len/Mac).
And for shits n giggles, weren’t all of those famous tunes built off of like 2-3 chords?! talk about minimalism. I’ll just call it all “analog techno” lol
Exactly! This is what I was implying in my comment elsewhere in this comment section. Artists have been using others for inspiration for centuries. Only in this present day of frivolous lawsuits can this be attempted and not laughed out of court. As I stated in my other comment, I believe the reason it's being attempted in this instance is because AI is an ambiguous, obscure, vague, gray area with unknown precedence and the Record Labels stand to gain millions.
Regardless of anything else, the RIAA is a relic that should just be disbanded. It's been running around the music scene for decades collecting money from other artists, more than AI ever have or will. Those who think "artists will earn more money without AI" are delusional. RIAA will earn more money, not artists. Artists will earn more money when RIAA is gone.
Two vultures are fighting over a carcass here. The music industry has been dead for a long time. Except for people like Taylor Swift, no one makes a fortune anymore. Smaller acts lose money on concerts. Only merch brings in a little cash. Hired musicians are fobbed off with small amounts. You can get entire catalogs of original music for a few cents. The labels that are suing here want to wrest money from Suno and Udio that they would never have earned through conventional means. Suno and Udio are trying to present themselves as the ones who are democratizing music as a medium. Even though they are trying to monopolize the music market. I'm watching the slaughter unfold completely untouched.
Yeah, I think much the same. One should perhaps look on AI music production as analogous to a Tin Pan Alley set of musicians/writers ready to pump out imitations of popular art forms.
I don't think a company like Suno or Udio has any chance or hope of monopolizing the market here. They've got, if we're extremely optimistic, maybe a few years during which can be market leaders in this AI generated music. More likely just several months. Then open source models will start popping up and proliferating that anyone can run on a good enough computer or run on a decent server. And if Suno and Udio are doing really well (not even anything close to monopolizing the music market), the other big players like Google and Microsoft will likely also try muscling themselves into that market and release their own versions. If they stay ahead and avoid any major PR blunders, they might remain a popular pick for an AI music generator for some time to come. I just don't see how they'd have any chance of getting a monopoly in this particular market.
Suno's response is telling. It complains about greedy record labels when in fact its AI program sets a precedent that no artist should be comfortable with. It learns your music so people can take your music and claim it's theirs. Even worse, it can be achieved with zero musical talent.
I'm a suno user since its first beta release and still just can produced some seconds music with weird vocal output. Since then Suno has been my playground to expressing my feelings. Bring my poems to be more alive. I have a bit of knowledge about big data too since i have daily job on that area too. And, i can say, suno has manipulative side by pointing that suno has trained like a kid's training or learning. No, it's not as simple as that. I can say the main problem here is money for that lawsuit matter. For tech person point of view, no lawsuit can stop what these developments will developed in the future, so people will always found a way to survive on every crisis they face as i read along the comments mostly users shootout their concerns about this suno will replace an artist in the future. It won't happen so soon but it could happen, so rather than afraid of something uncontrollable, just use tech at your best interest.
I think the arguments that Suno is using where they say that 'look at all the anti-competitive' things the big labels are doing, while it is true, it is irrelevant. All that is relevant is whether or not they are infringing on copyrights. 'Fair use' doesn't apply to for profit use, only for educational, etc., use. Creating something similar that is derived from existing work is a derivative work, and derivative works require consent from the copyright holder. So, my view is that if this software enables anyone to just push a button and create a competent song, the software had do download songs to learn from, and by that fact, it's a derivative work, and derivative works require consent of the copyright holder. "Transformative" is the word used in copyright law to avoid consent. Though a song maybe 'similar', is it transformative from the original? If so, no consent is required. But, it is a subjective term and it requires each song to be litigated and a jury decides if it is 'transformative'. So, are we going to flood the courts with thousands of lawsuits to figure out whether each song AI creates is transformative, or derivative requiring consent? This posits a nightmare scenario for the courts. Also, enabling untalented people to create music, is fine if they use it for non commercial means, but that is not what is happening, AI music is entering the market place. Untalented people are using AI to compete against artists who have spent their lives honing their craft. When they do that, they are diminishing the value of music. Otherwise, why are they arguing that the Big label's are being 'anti-competitive'?
Downloading a song to learn from does not mean it's derivative.. by any stretch of the imagination. ZERO of the original data is stored in the model. It quite literally learns via patterns and algorithms what a song is and how it is structured. Copyright law also throughout history has only applied to output, not to input. This precedent will most likely end up in AI's favor. Output can definitely be copyright claimed, but that would be on an individual basis. Otherwise Photoshop, Premiere Pro, DaVinci, etc... would be sued if a user used their software to produce copyrighted content. As far as defining transformative, you really can't get more transformative than distilling songs to weighted connections, running weighted connections through multiple million other weighted connections and using algorithms to end up at an output. IMO the value of music is not being diminished anymore than in the 2010's when home production and Soundcloud boomed. Good music will rise to the top, all the rest will fizzle, same as it's always been. Will it suck for artists initially? Probably. You could also make the same case for workers during the industrial revolution, painters when photography boomed, artists when digital illustration boomed.... Times and technology change, some things good, some things bad. The new generation will most likely combine both AI and real to produce music.
@@Hosea405 2 problems we're dealing with here: 1 - Although the algorithms seem to be able to learn theory and avoid copying songs in totality, their 'learning' style for voices, instruments, etc are more akin to 'stem separation' than 'learning a song' sampling _your own_ real instruments or _purchasing_ samples/parts from artists or people who record samples. In order to learn it has to, for example, separate a sample/audio of drums from a song, figure out what genre that belongs in, and then listen to prompts and roll the dice to figure out when it will be used/remixed. It can't copy an artist's voice or producer's tag exactly (as it has been able to many times) without storing those clips of audio and regurgitating it. Therefor while the songs/structure may be considered original to a large degree, the audio used to perform them is very much derivative. 2 - The problem for artists isn't really a matter of AI song generation existing; it's the fact that their voices, production, and playing can be used by other people without compensating them. If AI were done correctly, (although the progress would be slower/limited through people not wanting to be involved) each artist would need to be compensated for when their work is added into the neural network OR if traceable, when derivative audio is used to create a new song. Otherwise, AI music like SUNO and UDIO should be strictly non-commercial/ non-profit. There's a reason why VST's which Soundcloud musicians often use still require legal sources of audio/samples. This is like a super VST that illegally procured its samples. Which, again, wouldn't really be a huge problem if it were non-profit. If a new artist is on the rise, and I can use tags and other tricks to make a song in their style with their voice to compete with them, they should receive some sort of royalties for that, morally speaking. However, if the courts do get their shit together Record Labels and AI generators would both be nerfed, and artists would be compensated better than they are if at all. AI music is fine, but the data one trains on should be procured with permission if you're not using 'your own' audio sources and instruments. Meaning the AI companies would have to hire performers and find a way to use those performers to recreate the audio. And producers to teach AI how to add fx to the audio. etc. Rather than ripping and sampling it. Which is faster but shady.
With all due respect but I can't help but laugh at "Untalented people are using AI to compete against artists who have spent their lives honing their craft." since untalented people are being sold by the record labels for many years. There are many talented artists that are ignored by Record Labels in favor of an autotuned pretty face that cannot sing or write / produce any music. It takes a lot of money and a lot of people to make this completely untalented but beautiful person an established artist, but its ok since Record Companies have a lot of money and in the end the money go back to them anyway one way or another. So let's not pretend that hard-working artists honing their craft is our priority in the music industry. If every untalented person can create a song its ok, in the end of the day people will have to choose what they wanna listen to. We will always have to filter through the noise and find pieces of art that are worth of our attention. This was the same before A.I. and will still be the same after.
@@sheeeit7730 Labels only care about what will sell. That's true that that does not need as much talent for music as it does talent for making things that people like that many people could also do. (Kind of like youtubers that play the same game or do stupid things most people don't care to do.) However this software kind of rips off of artists who _do_ put effort into their craft as well. Meaning that statement wasn't false. The label companies screw them over as well, so I don't want the label companies to win... I'd rather see artists get compensated.
The algorithm isn't even learning, so the question is wrong. It doesn't understand what it is doing, but can instead mimic within a spotted pattern. If that pattern is [your songs] it will produce something so similar that you might get in trouble for copying _it_. But even if we took the question at face value, a big difference is in the effort to produce and the quantity. Even though AI currently only produces sub-mediocre similies, it is just about good enough to make it hard to spot at a glance and the amount of material that can be made by AI makes it very hard to find something that isn't slob. AI media thus destroys many markets without supplying any value.
Every song begins with a story, often drawn from the depths of the songwriter's experience and imagination. While lyrics form the backbone of a song, rooted in human consciousness, AI like Udio and Suno have the remarkable ability to interpret and amplify these human emotions. Creativity may start in the mind of the writer, but AI can enhance and expand that vision, offering new dimensions and perspectives to the creative process. The music industry should think about that.
An AI model doesn't store the input anywhere, and that will be easily proven. It just recognizes patterns very accurately. You just watch. Most of those legal experts will be very disappointed at the end, because they don't understand the tech.
@@edmundbardacha7211 If you’ve ever trained a big neural net on huge amounts of data, you know you need to store it locally on some fast storage media. I’ve trained neural nets for chess (Torch, Leela, Dragon, etc.) and had to keep terabytes of data around for retraining and experiments. Streaming would be terribly inefficient. Music data makes chess data look like nothing.
I'm a lyricist but not a musician nor a singer. I do not have the kind of money needed to have someone come on board and compose the music I like for my lyrics. I have used AI lyric to music and have made 3 great songs that sound nothing like any song I have ever heard. AI is all that some of us poor songwriters have to get tracks to go with our lyrics to be able to get tracks to get to the people we want to hear our songs and possibly decide to record it themselves
You simply cannot equate a machine learning or a neural network learning to a kid learning. That is misleading, and an oversimplification to make it sound completely innocent. Suno is and remains after all an Entreprise, and they made the machines learning not for the same motives kids learn. They made the machines learn to eventually drive profit, and as much as I hate the big major labels, I think the AI music startups willingly and knowingly took advantage of that gray area of what defines the public space vs to what extent copyrighting can expand (a music that you stream online may be copyrighted, but you can stream it regardless, or pay for it and make your friend listen to it on your phone. What if that friend is an AI company?). Moreover, the models running at Suno would have been completely useless hadn't they "used" music and trained on them. But I think that Suno's argument is valid in the sense that it won't be long before the same labels would have their own AIs generating music. So their ultimate goal is indeed to nullify any competition even on that front, but at the end of the day, all AI companies have secretely trained on the public internet data, very much knowing that this would definetly present a copyright problem.
While I do anticipate those record labels would build AI models and launch AI businesses themselves (like how Getty Images did that after filing the lawsuit against Stability AI), there is a cap on how how AI generation business, at least in the US: Machine generated works are NOT COPYRIGHTABLE. Those works would go into public domain, allowing other people to copy them without infringement allegations.
The way that neural networks learning to recognise patterns are not by downloading a subject, rather than statistically recognising a pattern. In simple language a neural network learns no “9” as a number as it is the most possible answer for the shape “9” in a text . It is not downloading the shape “9” , rather than this shape is the most possible outcome between all the other characters.
Which is why I personally don't agree with a lot of this legal stuff. Like suppose I learn from copyrighted physics textbook at school physics class that water will boil at 100C. Am I now obligated to pay money someone every time I say that water boils at 100C? No. But for some reason if ChatGPT were to say that water boils at 100C you would suddenly have 10 different textbook authors all claiming copyright to that knowledge. And I am not even talking about quoting the book because that would be clear infringement, I am talking about situation where you learn the fact and after that forget the exact text. And the same is with images or any other form of training data if AI looks at 1000 face photos and then concludes that face has 2 eyes, 2 ears, nose, mouth and then forgets the specific images it has seen then where is infringement? But for some reason you get like 1000 artists who all feel entitled to your money now because copyright. But if law was such that they would have to point out specific infringement on specific generated image they probably wouldn't win against you in court if your model was generalized enough.
@@test-rj2vl It's not really that different for text vs music (although if we were talking about direct quotes, there is more leeway for quoting a small portion of a book than for sampling a song's hook, say). If you ask ChatGPT to write a physics textbook, and in reality it's basically remixing 10 physics textbooks, and then everyone buys that book instead of the original 10 books, there's a potential issue there. I don't know how things will shake out, but from a rational, dispassionate standpoint, I think the fact that generative models are capable of spitting out fragments of their training material almost verbatim is pretty damning. If I were a lawyer, I would argue that portions of the copyrighted material fed into the system are almost certainly "encoded" into the model-maybe pointing out that training a model is essentially equivalent to jointly "compressing" all the input material as much as possible-and just because the people operating the model (which is a data file or set of files that is probably hundreds of gigabytes or more) don't *know* exactly how much or how well the input data is encoded into it, or just because you have to "prompt" the model with a piece of the input data (or something like it) in order to "decompress" the rest, doesn't mean it isn't in there.
@@davidg11235 Yea I understand your point, I was not talking about asking ChatGPT to spit out entire book. I agree that if in such book you can put your finger on exact sentence and say that it's stolen from another book that then it rightfully deserves copyright strike. However people would be complaining about copyrights even if ChatGPT gave out just one fact. Like if instead of asking it to write entire book, you ask at what temperature does water boil, it tells you at 100C and now every textbook author who has ever written about that fact feels entitled to money because their textbook mentioned.
@@test-rj2vl You can't copyright an individual scientific fact , any more than you can copyright a single musical note or chord. But if you explain that fact in a certain way in a book surrounded by other text and diagrams , or write a sequence of chords with a melody on top then you have created a copyrighted work.
This is absolutely absurd! If using music (or any content for that matter) for inspiration is worthy of a lawsuit, then there should be lawsuits everywhere. Every creator for centuries has been inspired by another creator or multiple creators. So, why aren't they being sued? Because it's absurd and the plaintiff would be laughed out of court. The reason it's being attempted in this instance is because AI is an ambiguous, obscure, vague, gray area with unknown precedence and the Record Labels stand to gain millions. And, the Record Labels are the bad guys, period! Greedy people who stifle innovation will always be the bad guys IMHO! And let's face it, the Record Labels don't give 2 💩about the artists, just their wallets. Sadly though, in this world of utterly corrupt court systems and judges, the Labels will likely pay off some judge to get their desired outcome and we will be the ones that suffer while the fat cats get fatter.
In my opinion, the premise to sustain the non-infringement argument is simply this: If you can prove you own every single note in every length or contrivance, then you would have a case to restrain this production of sound in those and any fashion demonstrated or performed. However, since you do not own every single note of any length or contrivance, then any creation of our system is merely a function of a reproduction of notes of varied length and contrivance. This is not duplication, its algorithmic contrivance based upon previous assemblage. Thus the machine is just projecting the next note just as ChatGpt does the next logical sentence, or in this case, the next probable note from the prompt that is the ignitor of the creation process, not that of any particular assembly of other songs. If a paragraph of ChatGpt is reputed as infringement, then it may constitute a sliver of hope to effectively authorize the cease and desist or infringement narrative. In conclusion, if you don’t own every word, you don’t own any notes, at all. Just because an actor has my nose, does not mean I owe the film company a dime when I’m on the cover of Rolling Stone for my Sci-fi film and my Grammy. It’s Napster 2.0
there is a difference between taking inspiration from a song as a human being, and a AI being trained on copyrighted material, that is infringement, stop treating AI like a person and like the algorithm that it is. AI should under no circumstances be viewed under the same context as ACTUAL working artists Ai regulation is far over due and its just making people like me and others who do creative work slowly become obsolete, which I would say is a pretty big negative for humanity
One problem is that there is a fundamental fiction that Suno and companies like it are perpetuating, which is that it makes a person a musician or a musical talent to write a prompt and click a few settings. They can call it "making music," but it is a completely different thing from real musical talent.
First thing that should not be overlooked: It would be literally impossible to train an AI like Sudo or Udio without actually storing and saving (and thus copying) the music data in some form or another. In order to train an AI, the training data needs to be input into the network thousands (or 10s or 100s of thousands) of times. It would have taken decades to train if it had to be streamed in real time. This is one reason why I think it's not unreasonable to consider Suno's training to be memorization (compressed storage) rather than "being inspired" or "imitating". AI is a meaningless buzzword that nobody serious in the industry should use. The field is called "Machine Learning" -- and the "learning" part of that term has a very clear definition (you can look it up) and it is fundamentally different from what we consider "a human learning, getting inspired or imitating", or something like that. Machine Learning applied in the way Suno does it is fundamentally a form a data compression. The most powerful forms of data compression are described and developed first and foremost as PREDICTORS of the data they are compressing and saving only the difference with their predictions, thus achieving a smaller data size than before (cite e.g. FLAC and Opus -- Opus is the audio compression UA-cam currently uses). There exists a form of compression called "lossy compression" that saves this difference signal with a tunable amount of precision, achieving unprecedented levels of compression with tunable fidelity. Certain compression models even contain generative algorithms just like Suno, except they were trained on styles of background noises, so that it can be generated instead of saved (cite mobile phone speech codecs). We don't say that an MP3 of copyrighted music is "just inspired" or "imitating" regardless of how low sound quality / bitrate it has been saved as. The case might be made that if we were to look at the size of Suno's stored data (algorithm plus weights), it might average out to 4 bytes per song or something similarly surprisingly little. This is not impossible, lossy data compression can be counter intuitive: incredible compression ratios can be achieved the larger the data set, because patterns are more likely to repeat, especially for lossy compression. Of course Suno's lossy copy does not seem to contain an exact copy of every song, but it has shown to be possible to extract strong likenesses of songs, just using Suno's interface to their data. Suno has admitted that it is their interface (not the data they store) that attempts to limit the amount of likeness to existing songs it outputs, meaning that the data that Suno possesses does in fact contain the information to produce much stronger likenesses, further indicating that it should be seen as a lossily compressed version of parts of the catalog. The above may be a rather technical explanation, but it refers to only existing state of the art technologies that are actually in use, industry standards, proven to work and function the way they are described. CONTRAST this to how instead it is Suno's claim that their "Artificial Intelligence" algorithm has been "learning" and "being inspired" just like a "little kid with a guitar" would, or something. They claim that a computer program is doing this explicitly human activity. This is a rather extraordinary claim that requires extraordinary proof. Despite the recent hype around AI, it remains to be seen whether the current generation of these algorithms actually possess any "intelligence", let alone other human abilities such as "being inspired". Progress seems to have halted a little, the bubble is bursting and over the entire field of "AI" the result is the same: outputs appear quite impressive to anyone not familiar with the field from which the data was stolen/used/"inspired by", but are invariably bland regurgitations and/or slightly wrong to anyone who is. (credentials: used to study machine learning)
It depends on how the models were trained. Seems like the music is the oddball here since maybe training data are similar to each other, and maybe incentivizing the algorithm to store enough information to "reconstruct" the training data. Similar things were mostly not observed for image generators and the biggest goal for any machine learning systems is generalization. If it's outputting training data too closely it's a failed system because it doesn't generalize well.
Compress storage?" You understand - having that much data would be impossible to really store on a single server yes out of the millions out that stores all this information?
@@adr2t what do you mean? To give an example, a single harddisk 4TB of storage already fits about 8.5 continuous 24/7 years of audio at even 128kbps MP3. And MP3 is kind of old, not state of the art. Even then the point of lossy storage is that it can go arbitrarily low, as long as it's fit for purpose. Which it is, as demonstrated by their plagiarism machine. Either way my point is about ways of looking at what these giant matrices of weights actually represent. On the one hand there is the extraordinary claim that these massive amounts of numbers represent "something" that has "learned to be inspired and imitate" something. On the other hand there is the idea that these numbers actually represent the songs in a very convoluted way that can lossily be queried for audio representations of that data and mash up combinations of the songs. To give my gut feeling, I've played with these services (Suno in particular) and there is just too much vaguely recognizable stuff coming out of it. I also see it only interpolate between existing things, never actually acting "inspired" and creating something new. I also worry that e.g. a cool lick or sound that I really dig is actually straight up similar to an existing song or artist I'd just be unaware of. And I just want to know. I think it might technically be possible to train a thing with full attribution of sources (you just have to slap another matrix onto it?), but it might be a bit slower. It would also be a bit of a legal nightmare to get off the ground, someone would need to recuperate artists for contributing X% to the weight activations of some song or another ...
@@metadaat5791 Na, 1 million MP3 songs at best, and there are a ton more than that on the internet alone let alone around the world. Let alone the amount of sounds , voices, instruments, that goes along with that. So sorry, no not even close even if you could compress it all. While I do agree it does come up with something similar in lots of ways, it also express that we listen to patterns of things as well. You wouldnt listen to random melodies either. If anything, it shows most music follows the same type of design if you ask me. This is why country music sounds similar even if it has different words to it. Thats the other thing to this, its melodies that are what we learn - not so much the music it self, but melodies that we reuse over and over again in different ways. Yes, there are law suits to this day over a few artiest using tiny melodies from each other when melodies was said not to be copy righted in the first place. English is no different in that same line up either when you break it down. Because its used to make a understanding of what we want to express. So are we going to start suing each other because we use English, numbers, etc now?
@@adr2t The problem is that the compositional process and the the 'performance' process are 2 different things. It can come up with 'the same song' but no body would care if it sounded like a band in that specific genre you haven't really heard playing that song before. If you hand bands the same lyrics, chords, arrangement, descriptions of parts (ex: kind of quiet chugging rhythm guitar in the background of the verse) , and overall vibe, they would still come up with largely distinguishable recordings of the same song. The voices, specific instruments, specific production techniques etc... Suno and Udio are ripping off of and remixing the performances of real musicians/producers to create performances of their newly composed songs. That's the only real problem here. It should always sound like enough of a combination of different people within the genre to not be able to pinpoint a single voice or producer tag, or any likeness of a specific person's work. And it can. Which means it's engine is doing something more similar to advanced stem separation and using that to 'perform' its compositions.
As a creator who is fascinated and impressed by Suno as another tool, among my collection of various synthesizers and sample libraries, I tend to be drawn into the arguments that the output is also “stolen” or “cheating”. I thought of an interesting analogy. A bottled water company stole a bunch of water and started mixing it in with non-stolen water(the ratio is nearly impossible to truly know), and then I buy a bottle of the water. I then use this water to make a batch of soup that I’m very proud of and wish to share or even sell. Is the soup stolen? Does this impact my level of skill or talent that it took to make the soup? Yeah, I could make the soup with water that I know wasn’t stolen but I really prefer this water and it does make my soup better and I paid good money for it. This is probably pretty silly and not a perfect example but I thought it was fun, so take it for what it is. I’m very much interested in having long form civil discussions about these things so chat me up if you’d like!
What do you do when platforms like spotify makes AI from its users, then makes fake "artists" in a bunch of different names and pushes this AI music out under their own platform?
That is the problem. This is what people need protecting from, not the person who just finished work and wants to spend a few hours prompting an algorithm and trying to turn those written lyrics into a song people can listen to.
Once again, ❤ your take on what’s going on in the industry. I am one of those Suno users and fortunate to be growing as a new music creator. You consistently approach topics with a healthy skepticism and the emphasis being on serving the needs of your client.
Appreciate the thoughtful response. Please join us for our UA-cam Livestream tomorrow.. We host every Wednesday at 5 pm PST. We would love for you to engage there.
I'm a programmer, and the idea that "AI learns the same way a child learns" is a huge misconception. Not only it is an oversimplification - because we still don't fully understand how our brain works and especially the physics behind it, it also doesn't consider things such as our perception, the ability to transform information from completely different sources (audio, visual, memories, feelings etc.), and high level conceptual thinking. I am not good at law, but as far as I understand it, neural network "learning" of copyrighted material would be completely illegal, if people at court knew exactly how they work. Here is an example. Let's say there is some copyrighted sample pack that has a license which prohibits using it in your own sample packs. But you really liked that "Drum 01.wav" sample, and you want something similar in your pack. What would you do, as a human? You listen to it and analyze it, like "there is a very punchy part in high frequencies, and then there is a long bass tail, and it's recorded in a large hall". That's the learning part. Then, you go and record your own sounds, and make something very similar, and this is completely legal. How an AI would work here? I need to simplify the explanation so it doesn't take 1 hour. Basically, we have a copyrighted "Drum 01.wav" file. ".wav" files are basically a sequence of numbers which represent the sound wave. Let's say it is a sequence of the following numbers: "1 2 3 4 5". So the neural network takes this file (as a "training data"), and stores the patterns inside itself: that after "1" there is usually "2", after "2" there is usually "3", and so on (it literally kind of adds up, you have something like pairs of transitions and how often they exist, for instance transition 1->2 exists 1 time, transition 1->3 exists 0 times, and so on). Yes, the file is no longer stored in it's original form, it is now stored in "weights" thanks to some math transformations, but it is still the same copyrighted data, just like a ".mp3" format could be used to store this copyrighted sample instead of ".wav". So now, when you prompt it "Drum 01.wav", neural network gives you your "1 2 3 4 5", but now without any copyright obligations, or at least many people believe so. To make it even more confusing, you can add 1 more or a million more copyrighted samples to the training data, and all these patterns (which are basically very small parts of the copyrighted material) from all these files will add up, and it will be extremely hard to tell which exact samples are used in a training data, though definitely possible, unless such outputs are specifically excluded for users. So, as you can see, this AI "learning" has really nothing to do with human learning, even though it may seem so from a distance, and from my understanding of the law, it just can't be legal by design. However, I think that this technology has a lot of potential, and that eventually we will be able to create such a system that will compensate all artists whose data was used for training or even to get certain outputs. Sorry for bad english.
It's misleading to say that machine learning algorithms are inherently "collage" and "copyright infringement", which is what I took away from your argument. Simply put: it depends on how the data was processed and what information was extracted. Human has better abstraction capability but that doesn't mean anything machine learning on copyrighted material is automatically copyright infringement, despite the abstraction could be simpler. In fact: I can argue that improved abstraction capability is one of the biggest factors that push generative AI to the mainstream right now. Not defending any particular product here (in fact I believe that if someone has proven to break the law, they should take responsibility) but it's what it is.
@@voidmain7902 It actually does, again, specifically because AIs deconstruct copyrighted files, which is prohibited by license. And while it may depend on how the data is extracted, again, this data is EVERYTHING for neural networks, you can't do anything meaningful without high quality traning data (which is usually copyrighted) - unlike humans
You are not a programmer.... sorry.... the fact you had to go to the whole idea of "how our brain works and especially the physics behind it" means you dont understand how AI in general works
Thank you for the support. Please join us for our UA-cam Livestream tomorrow.. We host every Wednesday at 5 pm PST. We would love for you to engage there.
Ai is scanning every pixel of your picture so everyone can prompt and generate something similar without having to buy your picture. That is what they are doing. Trying to put you out of business so later they can impose the "Universal Basic Rent." Which is their ultimate goal.
Serious question... I go to a studio guy who plays keyboards, guitar maybe he sings. His buddy plays drums and he hires a sax player. I'm paying for his services. I give him my original lyrics and say I need a song for my movie... "Imagine the band Chicago meets The Cars". Isn't he learning by listening to their music?
Suno is as much a new instrument as is a player piano, Humans are not hitting the ivorys and any one to take the bench and pretend is akin to suno users today
@ It’s both the composer, and all the instrumentalist, my friend, well as vocalists, including harmony vox tracks, studio, tracking, mixing including effects, mastering and cover art, lol If you didn’t touch the keys on a computer to enter a prompt, suno wouldn’t compose anything either. I don’t agree with your analogy
@gitarman666 Yes, exactly. It writes the music for you. You don't write the music. And like you say, it includes everything in the way of performance and production.
I thought about not commenting on this, cause i dont want to expose my hand. I have worked with all of these labels multiple times, one of them still handles my publishing and distribution. they are absolutely terrified of the competition. The average user isn't even the concern...the concern is the people with a really deep knowledge of music theory using these tools. Some of my friends (I cant say who) that are multi platinum selling producers, have taken time to learn these platforms on a deep level...and the music they are getting out of them is mind boggling. truly next level stuff. This is why the labels are so afraid.
Once people start understanding that these AI tools are so deep to the point you can tell it what key to play in, what chord sequence to use, etc...the labels are kinda screwed.
As songwriter channel I'm using SUNO intensely. It's a great way to make my lyrics come alive. Saving immense pre-production cost for demos. The lyrics are the most problematic part of AI's. That's why I never use SUNO lyrics. BUT I can generate great music for my lyrics in high quality! That's just awesome....
Does not us the users train AI? Biggest reason I use is I hate the music.corp they take from us, sell songs they have stolen, we never have the money to fight them, I rarely listened to music till I could write MY own lyrics and AI puts music and vocals. I have a note book full of poems and lyrics from over my 63 years. Tried to sell one in 1979, person just took it
Users of AI services are customers. They are the ones being trained, they are not training LLMs from the armchair in the living room. You could pay actual musicians to put your words to music. Imagine that. Supporting people in your own community.
this happened to me too back in 1998, I was heartbroken, it took me out for a while and took me until 2013 to start writing again, I am thankful for Suno giving me that voice back
I agree 100% with the argument that it is not a copyright issue-not even a derivative work. Fair use enters for the download aspect in that it transforms it into statistical insights-not even music. It has no impact on the commercial value of the works. I suspect it even takes segments rather than entire pieces for legal as well as technical reasons. A style, a musical vocabulary, is not copyrightable. But I also believe that creators deserve compensation; they're not just teaching an individual, they're providing that training to myriad users of that NN. We really need new legislation here. It could leverage the existing performance-rights organizations. The issue of the safeguards, needs to be viewed from the lens of generating vast numbers of outputs. This increases the chances of creating works that might be deemed similar. It does sound from the complaints that these safeguards aren't adequate, at least from a protect-from-legal-jeopardy viewpoint. I have to wonder why they don't license content-ID tech! BTW, from an old AI guy-you're doing very well with the AI stuff. And I'm loving the sort of joy you bring to analyzing the legal arguments here!
Clearly you haven't watched the video, seeing they were able to recreate exactly Johnnie B Goode songs just by asking for a genre of music out of the software and another 1000 songs.
Well the AI is trained on final songs and they are wet, and you would really want access to stems and dry samples. Artists have those and they are not public. They can make a deal, and in the process get an AI that works in stems and is better for them to use. Also labeling is critical and Artists have the understanding required to better categorize, a paid job. The difference between GPT and ChatGPT is that ChatGPU has humans in the learning process, and giving feedback on outputs.
@@sredz8702 Exact? No, not quite what it said. With all the myriad covers of that song, it can basically create another cover, drawing on all the covers. I agree it's an issue; that's why I suggested licensing content-ID to detect and reject such output. Learning a piece and performing it does become a copyright issue, but that's a special case, that the safeguards are supposed to prevent. The main issues are: does training involve copying, and the copyright implications of that, and does a model contain a copy, which is particularly a question if the model is distributed. Your point relates instead to the output-and if you sat down to write a song and Johnnie B Goode came out, b/c you'd heard it but forgotten, that raises the same issues, whether AI is in the loop or not.
@@abram730 But it depends on what you're trying to do. Training on mastered final mix will produce things that sound like mastered final mixes. That's ideal for some audiences. If you're using AI as a starting point or element of your creative process, yes, I'd want stems, and that's a great way to get some revenue into artists hands. I'm not sure where you're going with your point about reinforcement learning with human feedback (RLHF)? Where do you see injecting humans into the learning process? Adding humans, even for labeling, is expensive. Natural language prompting does require some form of labeling; I suspect they build on recommendation algorithm data to minimize it?
If y’all are surprised at how Ai “reads” data to make a piece of music, wait until you hear about & understand how ComfyUI works with “input video” as source material. Fun stuff for sure 🎉
I learned how to play guitar and write songs using my own intelligence by studying recorded material. And now all my playing is an original rearranged imitation. And as a result, a lot of the stuff I create resembles to a certain degree the style and methods of my influences.
We found that people will always listen to the music that they like, no matter the competition, AI or otherwise. At least we have choices and variety, in a world where "monopoly" seems to be the real aim. Stay "seriously" awesome 😁
I am a musician and I use A.I. as a tool. A.I. has to use what already exists in order to do what it does. That's not copyright infringement. Using A.I. to generate tracks is no different than an aspiring person listening to their favorite artists from their favorite genre to create their own tracks within that same genre. That's the way it is. My favorite genre is blackened thrash metal. There are thousands of bands in that genre who were influenced by other artists in that same genre. We learn from each other. We pick up on different ways of doing things. We have to listen to each others music in order to do that. However, my bands tracks are not copies of anyone elses tracks. Are there similarities? Of course there are! It's the same damn genre! The mainstream record labels are full of shit. Plain and simple. They feel threatened by A.I. generated music because they aren't getting a cut. Independent artists are able to sidestep them all together and they don't like that, because they can't control the artist and get a cut of their profits.
Part of my research involved interpolating across the latent space of a trained model. If you get the exact point in the latent space of a particular data sample of a generative model, it will reproduce almost exactly the original (assuming the model is trained well). Then, knowing that, you can interpolate between two sample points to create some sort of mix between the two points, by walking smartly through the latent space. Furthermore, you can get attributes from a group of data points in the latent space and calculate its vector, and merge attributes on a data point to generate "new" content, from a mixture of the previous samples. The fact that these guys are trying to claim that these are not samples is asinine, because this is almost exactly what sampling is doing. Speaking as an ex-PhD student in trusted AI that spent years studying this.
If you really know what you are talking about you'd realize that latent space is just a fancy way to represent data, and it largely has nothing to do with what the neural network is doing to it. Sure the neural nets can do something as simple as biased randomization or linear interpolation, and I agree that this in particular would be some form of sampling, but for a lot of other implementations this is simply not what they're doing as they achieve better generalization. This is the same thing as arguing that me drawing on an FFT plot of some samples so it's no longer recognizable, or me attempt to redraw some samples from memory by hand in PCM or FFT domain as some form of "sampling", and put these in the same bucket as just pass the entire sample through a low pass filter.
they're lossy "samples" though, due to the nature of the neural network. it acts as a very specialised lossy compressor in the latent space of its training distribution.
@@erkinalp Look into CCN's specifically as a tokenized CCN is what is believed to be used here. It creates features, or generalizations at a low level and complex abstractions at a high level. The intelligence is emergent, as it isn't specifically taught, just needed to function. One could imagine latent spaces, but at a low level you have waveform patterns like a synthesizer that works in both directions. The imaginary latent space only exist in a very abstract way, like a pattern of presets for a synthesizer, and patterns to play them in. The latent space being a conceptual one made out of instructions to forge a song using no direct samples of the song, and with the labels that describe the space. The AI simply learns to stay away from the location of the song painting in the spaces between. Understandings happen even deeper in the net. How would one combine Asian throat singing with death metal? How would one make a Honky-tonk dancehall fusion song? It needs to create a latent space between two space that are far apart and don't at all match. That requires some deep understandings and lots of guesses in terms of intent.
AI is a separate form of intelligence and it has learned from the whole of music the way birds learn from other birds how to sing. Artists have always drawn inspiration from other artists work. In music college we learnt from some of the greatest tracks. AI has learnt what music is. From here the whole of humanity will change in relationship to AI.
So after AI companies have created 500 billion songs, they will attempt to sue real songwriters claiming new music closely matches one of 500 billion melodies they created w a push of a button... this is heading off a cliff.
@@royaltyfreemusiccollective8662 They have not as yet lost the case ! And if given a fair chance they won't! Every song written in the last decade can likely be traced to music that has been used before if looked at closely! That's why they can classify music into genre's! The big question here is should anyone be allowed to own and control music? I say no, it should belong to everyone
One sentence: You can never copyright how one learns. Don't see why ppl think this is a legal issue. What is the legal issue is if anything is "exact" copies, and if the lawsuit isn't abusing the software to "make it" exact in copy. Lawsuits started right after importing went public. They only have only 2 weakness: 1. baby-level AI can leak mistakes from it's learning if it's too new to go public (Stable Diffusion had this problem until it didn't, and other's like it. And they're still up and active - music is no different), and the abuse of importing and editing features. This is also why if someone wants to profit off of a song they designed with AI, and intentionally make it similar to real song, then it's the user's fault, not the tool. And I agree with licensing, but not to this over-reaching scenario. This would only stiffle AI, and give innovation to simply a different country because of people's litigious greed in this country. Artist won't win if labels win the AI argument, since they'll make their own and simply copyright all melodies in, like, a day, and put it all on Spotify/Apple. Then, there you go, the profit loss with industry artists and indie market is resolved for them, but a fail for everyone else, but this time, with no chance of a lawsuit other than having a supreme court monopoly lawsuit like Microsoft faced. But companies maximize until they're sued to this level to reassure profits. Capitalism is simply an amoral system repeating itself in a new cycle. 23:00 That youtube has the incentive and intention already stated he wanted to destroy AI music and does so with smarmy behavior (intentionality = bias). So he could do anything malicious, and so can the label, to fabricate the issue. If you actually use the AI like Suno, it's not possible to replicate unless you do a lot of chopping and editing and spend a lot of money to eventually render what you want, since it's based on how you make it after editing it so many times. I'm not familiar with Udio, and it probably may be more compromised than Suno. But for Suno, it requires intention and a lot of work to have a suitable lawsuit, and even then, they may have to replicate it on the software in court to reassure they aren't abusing the court system, unless they have evidence from start to finish, in video and time of recording, to prove they are not presenting a false evidence. AI companies have the money to generate so many songs with the editing/import features to alter not only a voice but also melodies, and if Suno has the proper format in place unless it's too freely up to the AI, they could provide evidence of ill-intent later on in the lawsuit.
Such an interesting topic, very eager to see an update. A couple thoughts that came up while listening. First, there's a lot of misconception and unknown aspects of how neural networks and machine learning actually happens, but one thing that is pretty consistently emphasized is that generative models are able to generate because they become execeptionally capable of predicting the logical next step in a process or piece. So, while on the surface, it may appear that prompting the model to generate you "1950s rock and roll, 12 bar blues..." etc (and providing it the literal lyrics of Johnny B Goode) causes a copy or derivation of the actual original Johnny B Goode to be produced, in reality the structure and flow of thar song follows certain stylistic rules and trends which the model can predict and then follow. It's similar to how you can find videos of three chord songs and how they can easily transition one into the next of seemingly disparate songs which actually rely on the same fundamental (and predictable) chord progression. Which makes me also think about the section in the answer where they wrote that there are literally hundreds of cover versions of Johnny B Goode. I don't know if there is case law about this, I suspect that there may be, but I'm not certain, but if the model was trained on 7 covers of Johnny B Goode, is there still an argument of copyright violation as it pertains to the original copyright holder of the song? The covers are not infringements, so if they are the training data source, is there any claim of copyright to be saught by the original? That may be what the answer was trying to alude to, that if I show the model 17 covers of Johnny B Goode, I don't need to have ever copied or shown it the original protected song, it will still "know" that song structurally, stylistically, and highly accurately.
I think it's more of a sound simulation algorithm. It's not like comparing two songs that sound similar, it's like comparing guitar amplifiers. Crate and Marshall both create a rock distortion. There's no way to copyright "Crunch", is there?
The worry regarding AI systems replicating the work of current artists and the associated need for fair payment is understandable. Nevertheless, it's crucial to recognize that these AI systems do not simply replicate existing creations. Instead, they identify patterns from their training data to generate new material informed by those patterns. Although the output may draw from previous works, it doesn't constitute a literal reproduction. AI systems can be viewed as taking cues from existing creations, much as humans do. They construct new content from learned patterns but do not possess the conscious intent and creative flair that people have. Their functioning relies on algorithms and patterns rather than on artistic decisions. To sum up, AI models can generate unique and original content by employing learned patterns and adding variations. While ethical issues surround the concepts of inspiration and resemblance to existing work, it is essential to distinguish between actual copying and the generation of new content derived from learned patterns.
Exactly. And in the process they subvert the filters looking for violations. The program isn't creating it, so they technically don't need to check added lyrics.
I'm a published musician, programmer and very experienced with AI. I've even built my own neural net from scratch just to understand how they work. The question becomes how exactly do the neural nets of Suno and Udio work. All AI experts seem to agree that we have little understanding of how a neural net works under the hood. But the fact that you can get a producer tag that sounds exactly like the copyrighted audio is proof that at some level it's storing those samples. Literally proof, not up for debate. When you get to models with billions of parameters and many levels deep, it's very possible that a single neuron might be storing the actual wav file in some form, just because they don't need to doesn't mean they aren't. If you compare the quality of Stable Audio to Udio it's clear that Udio has much better sound quality. Well, it just so happens that Stable Audio is open and transparent about their NN inner workings while Udio is closed and proprietary. It's pretty obvious to me that this is because Suno and Udio knew that what they were doing wasn't 100% above board. And all of this talk about the various mechanisms in place to prevent copyright abuse is just nonsense, it's very superficial and not even effective as it's easy to circumvent. No, Suno and Udio exploited the fact that there is no precedent set here yet and knew they could get away with making millions of dollars off of other people's creative works until something is established - which could be a long time to never. But, the truth is that neural nets emulate the grey matter of the brain, but they are no-where near an actual brain - and to suggest they are "learning" in a way similar to how a human does is gaslighting in hopes they can continue stealing millions from the actual creatives until someone points that out (and wins in court). I'm perfectly fine with saying that an AI company MUST have licensed the rights of all the training data used for that model. It would make the quality much worse for now, but it'll get better, and it wouldn't just be these few companies that have such a dominance - it would allow for better competition - aka capitalism.
Look into CCN's specifically as a tokenized CCN is what is believed to be used here. It creates features, or generalizations at a low level and complex abstractions at a high level. The intelligence is emergent, as it isn't specifically taught, just needed to function. One could imagine latent spaces, but at a low level you have waveform patterns like a synthesizer that works in both directions. The imaginary latent space only exist in a very abstract way, like a pattern of presets for a synthesizer, and patterns to play them it in. The latent space being a conceptual one made out of instructions to forge a song using no direct samples of the song, and with the labels that describe the space. The AI simply learns to stay away from the location of the song painting in the spaces between. Understandings happen even deeper in the net. How would one combine Asian throat singing with death metal? How would one make a Honky-tonk dancehall fusion song? Producer tags likely aren't flagged by copywrite checking software, likely used in training. Thus the AI does not learn that forging them is bad. If a person can do impersonations, does that mean that they have made an illegal copy in their brain?
"a single neuron might be storing the actual wav file in some form" 😂My man, you have no idea how a neural net works. Neurons don’t store data like files on a hard drive. They don’t keep copies of the training data. Instead, the entire network learns a representation of the data through the adjustment of weights and biases across all neurons.This means that the knowledge or patterns learned by the network are spread across many neurons and connections. There’s no single neuron that holds an entire piece of data (like a WAV file); instead, the information is encoded in a distributed manner across the network.
@@rogue_bard I said "might" - I don't know, you don't know. I know that if the output is a bitmap that sounds EXACTLY like famous producer tags then either they're using so many params that it's overfitting or they're doing something else that we're unaware of. It might be overfitting and then as abram730 said, being missed by their copyright checking systems. Bottom line is that it's proprietary software, so all we know is that in some cases, it's replicating the exact input in the output. But what it is is absolute proof they used songs with those producer tags in the training data - then the question becomes, did they have permission to do so, or did they just do it in hopes to get away with it.
@@EarthAaron There’s no "might" about it-this simply isn't how neural networks operate. A neural net isn't some magical black box; its workings are well-understood by those with the relevant knowledge, even if that might not include everyone. When you claim that Suno or Udio sounds exactly like a famous producer, it suggests either you're listening to genres that heavily rely on synthesized and over-processed sounds, or you may not have a highly discerning ear for sound quality. For genres that prioritize more natural sounds, the audio quality from Suno or Udio is GARBAGE, more often than not. This is primarily because the system was trained on final mixes (I presume), not individual stems from each instrument, making it extremely challenging to achieve high-quality sound. If artists provided original stems from their recordings, the sound quality could improve dramatically. This could also pave the way for a quid pro quo arrangement, as many have suggested-where Suno/Udio pays artists commissions similar to what streaming platforms like Spotify offer. Moreover, if you believe that success in music is solely tied to the quality of the songs, I'd argue that’s a misconception. In my experience, success is largely driven by marketing, though, of course, the music itself needs to be good (whatever that may mean in certain genres).
@@rogue_bard actually, since you don't know how these proprietary models work, there "might" be a million monkeys banging on keyboards that's producing the final output. The truth is that they're probably using a transform-based model and tokenizing the spectrogram of audio chunks. Now what's the difference between an audio spectrogram and a wav file? A wav file stores the sound pressure over time, a spectrogram stores the frequency over time; so, yes, I should've said spectrogram instead of wav. But, either way, they "might" be doing that or neither as I'm merely speculating. And apparently you're not aware of the producer tag debacle that Suno is mixed up in -- generating producer tags as unique as a fingerprint in full clarity for anyone to verify. Maybe you should look into it. And the fact that neither platforms produce hybrid sounds very well suggests that it's highly overfit for popular music and is less like GPT in that it's probably not just guessing the next token. I haven't seen anything to suggest that Suno/Udio are considering paying artists for using their content as training data. This is the whole issue. The facts are that they've both created magical black boxes, aka proprietary software, with evidence that they've used copyrighted content to train their models and even though we don't know the inner working of their software, we do know that the output can easily create music that debatably infringes copyright.
I think the simple solution is to just not allow Ai generated music to be for sale. I like that a lot of platforms are banning the sale of Ai generated music. On the other hand, I personally like using platforms like SUNO to get inspiration for songs that I can then take elements from and build my own music. But the truth is, the Ai generated music still sucks for the most part and I don’t see it having the ability to generate music with the same high fidelity as commercially released music for the foreseeable future. In the meantime, let people play around and get inspired with it. But as an artist I don’t see the value of profiting off crappy Ai generated music. It should be seen as another tool for songwriters who need a fresh perspective and need some inspiration. Ironically, the whole record industry basically has been making derivative works and outright stealing from other artists for decades. You see it all the time in pop music where producers just study other songs and try to make something similar. Ai is no different.
I made a tutorial about how to create sample replays for the funky drummer. I posted a clip to UA-cam that contained no audio from the original recording and triggered the content ID. Just wanted to add that experience to the discussion. I appreciate your videos and unique insight as an artist and attorney. Thanks for you content!
There are truckloads of free music available that can be used to train AI models. I also think that platforms like Suno could untrain or instruct the model to avoid using specific songs.
Lots people defending suno , because they are fking lazy to learn how to make songs or sing giving prompts and calling themself artists, and people saying it's like kid who is learning music by listening (machine learning) get f outa here machine learning and Ai is different thing and also these big companies like suno and big label companies they both are evil i reject both; but products like suno make things more worse for both big and small artists who want to make living out of it
I hope Suno and Udio win...my mom used to write a lot of poems over the years. Some were playful, some were deep, some could have been made into songs. She died at the end of 2022, so there has been a void in my life. even though she used to annoy me calling all the time...I just let her do it knowing I'd miss those calls someday, but to the point, I'd like to use AI to hear her read her poems using voice samples I have, and I'd like to turn some of her songs into songs. Plus I kind of have that gift of poem and lyrics writing...it's so cool to see something I wrote being put to music. Now, do I claim I'm a musician? No, author would be more like it, but an author is something....though I do have FL Studio and dabble with samples and VST's, but even then I considered myself an editor, again....still something creative.
@@marksmusicCC hard to express, but it wouldn't do my mom justice keeping her creative works locked away, and not appreciated by others. I'd have to clear it with my brother first anyway, but like I said, she gave me the skill of writing. Most of the songs I made in AI were written by me, though I have been lazy a few times and prompted some.
I think that would be an amazing act of LOVE & RESPECT to honor your moms creative spirit through recreating her works in an collaborative & augmentative way. the tech is definitely catching up so that is already possible. May also open you up to releasing the heavy weight of your loss & bring some comfort in knowing her works not only impacted you but also may have uplifting resonance with others. & don’t forget to Have fun!!!
Meanwhile every still living artist having other people using AI create new tracks just like theirs without permission being paid to them instead be like "....BRUH...."
@@1wibble230 No, cause every genre has it's own formula, and then the artist produces their own variation within the formula. With AI you can mix match different genres of music. There is no limit to the creativity, even though they're not "Musicians" per say, they are authoring unique tracks. You could have it do a Heavy Metal song based on a Mideastern scale, and then maybe add some EDM, or orchestral elements to it. Just cause AI learned from copywritten material, just like every other living artist did, does not mean it's out putting copies, or remixed samples.
Ok. Let me put it this way: what does “original music” mean? Because, if you try to define it, you will have very big problems. Every musician was inspired by others one way or the other, so when does “original music” start and where does it end?
I knew this was coming and I caught him before it happened. A few months before the lawsuit was even announced, I downloaded an interview with the CEO of Suno which has now magically disappeared off of the internet where he outright lies about how he trains his models. He said he got permission from artists who own the works to use as training data. I called bullshit and pumped out a song that closely resembles a massive pop star using a combination of ChatGPT, Claude and Suno. The lyrics were 100% her style and I had that track that I generated with suno deeply analysed by all the other AI's on completely fresh data to remove bias and "memory" and they all "assumed" that the track was potentially an unreleased song by this very artist. Meaning, with enough precision you can create a song by any artist, even if the vocalist doesn't sound the same. Proving, that suno does "mimic" the exact samples, synths and "patterns" of other artists. Which is moving from a grey area, to black hat one. My thoughts? Suno should pay up - Yes its cool that anyone can now make music, but based on other artists decades of skills and knowledge with no work involved? Completely unfair.... Go learn some skills, and sit alone in the studio for 10 years if you wanna use my work or pay me and most importantly, credit me, because you didn't make it, you generated it. Only then would I be happy about this "innovation".
I'd be interested in that interview - is there anywhere to view it? Also, I mostly agree with your points. Although I do think AI music is here to stay, but they should at least use a proper licensing protocol to train their models. These companies like Suno and Udio are trying to get away with mass theft imo
Unfair you say? While the rest of the music industry is indeed completely based on merit and is 100% fair? I think you might have stayed in that studio a little too long. Then go get a piece by Bach and release it as your own via a distributor. And sit on it. Don't do any promotion whatsoever, don't spend a single Dollar on advertising. Let's see how many streams / views etc you will get after a couple months or a year. Then tell me how it's fair that a musical piece written by a Genius sits there with 20 views while Taylor Swift's new single has billions. What are you really protecting here? Music? or the music business?
@@sheeeit7730 Your assumptions are hilarious, read my comment again. I am protecting myself and my fellow artists. I don't work with labels, I agree the music industry business tycoons are not protecting artists, they're protecting their profits. However, for your information, I went and learned how to release tracks and market them myself so I don't have to rely on a label. These days anyone can blow up a track with social media platforms like TikTok. If your track is good, its good. Simple as that. If its not a good track and the world agree's then that gives you the opportunity to become a better producer doesn't it? Millions of people like Taylor Swifts tracks, I am not a fan of hers but it doesn't take a genius to notice that she consistently pumps out quality and you're also forgetting about the producers that produce the instrumental of her tracks and the vocals and some of those producers I know personally and they worked hard, for years to become eligible and qualified to produce a track for Taylor Swift, simply because they're good. Take a look at Max Martin and you'll realise its him behind a lot of the biggest pop tracks ever, an absolute musical genius. That's how we grow as humans. If you're going to sit there and complain that your track has no views and do nothing about it, then that's on you. I don't think you understand how the music industry operates and how big artists with millions of streams barely get any of the royalties anyway. I know a lot of big labels that take 75% of royalties from the track. This would be fair if some of the money gained from the labels royalties was used to actually market the track but a lot of the time, their overheads are paid out from the artists royalties, and then after that, the artist gets paid. So again, read my comment carefully, I am agains't Suno here. If you want to be a producer then learn how to make good tracks and spend that time doing so but its not fair to just pay a subscription to Suno and automatically generate tracks that you can release on spotify etc when the generated track was based on the decades of hard work by the producers of the tracks that the track is made from, which is what suno has obviously done here. Did you make the kick? did you write the melody? Did you write the lyrics? Did you spend hours in a daw mixing and mastering it? Or did you generate a track on Suno that recreates elements of actual tracks that are out there?Its basically sampling, because you didn't make it. Its like if you listened to a track on spotify and then you went and created that exact same tracks elements for a new track. That would take hours but is still copyright, AI can just listen to a bass pattern, or a kick from a track and recreate that exact same kick drum and bass pattern instantly. Its still copying. Need another example? What if you saw an artists painting but you knew exactly how to paint the exact same thing and it looks identical to any eye. Are you saying you're not stealing because you just remade it? AI can also do that instantly.
@@EarthAaron Exactly, I would be 100% okay with it if they trained their models on tracks that they either made themselves, or hired producers to make specifically for training the models or properly licensing the tracks that they don't own for training the models under the obvious provision that they can't just make the same shit, exactly the same. What happens when a user accidentally generates a track that sounds virtually identical to an original release but because there are so many tracks out there, they didn't even know and then released it on spotify themselves and racked up millions of streams because it already sounds like a well known, good sounding track that is already out there with millions of streams. Who is liable for that, the user, suno, or both?
For such a detailed review. If it helps you - the creators of suno are trying to mislead the court and ignorant people. As experts say, the model contains complete copies of works, but only in the form of weights. It's just like a new compression method. And with certain input prompts and random seed, they are able to produce exact copies of the works on which they are trained. If you compare the training of a child, this is standard manipulation. A child does not have the ability to find and create works from a base of weights in seconds, producing a sound wave. In general, it looks very dirty, especially the forbidden example with an elderly person. It remains to add poor hungry children from Africa who generate songs.
I do hope Suno AI wins. I personally use Suno AI. These are my points. 1. I lost my voice. Suno Ai gave me back my voice. 2. I have learned different song structures, genres, beats, etc instrumental instrumental
3. Learned how to write lyrics of different genres, how to build to make a song. 4. My lyrics are my memories. Big Music just wants to continue to monopolize. The big company is upset that Suno AI is not controlled by them. They would use the technology if it was exclusive to them.
Me too, I was scared to comment but...I can write lyrics for days, but I do not have the "look" that record labels want, not only that I'm goofy in front of a live camera and I will stutter in front of a crowd and fall all over my lyrics, Suno gave me the voice I needed. I did ghostwriting back in the late 90s, I will never do that again, it's a very sad place, I am happier now barely making a penny off my stuff than I was when people were paying for my lyrics back then
@YTTraveler777 You are not an artist. You couldn't think of something more creative to say? AI is a tool. Do you play all the instruments? No! You can write your lyrics. You don't have to pay a band to record your songs. You can generate the instrumentals and vocals. Your lyrics are not just sitting on paper. It gives bands, vocalists a chance to hear more songs quickly to find the ones they want to perform.
Thank you for the support. Please join us for our UA-cam Livestream tomorrow.. We host every Wednesday at 5 pm PST. We would love for you to engage there.
What's funny is that some AI "artist" are acting as if they are gonna become super stars, famous or take control of the music industry if the record labels lose the case. Nothing will change.
I felt the title was a bit clickbaity. So I had ChatGPT pull this episode apart for you all The title "Major Music AI Lawsuit Update: Suno's Shocking Admission" appears to be somewhat clickbait, as it implies that Suno made a dramatic or unexpected confession. However, the content of the transcription reveals that Suno's so-called "shocking admission" isn't as surprising as the title suggests. The main point is that Suno acknowledged that their AI models trained on data from the open internet, which may include copyrighted material owned by major record labels. Suno's defense, however, revolves around the argument that training AI on publicly available data (including copyrighted material) is akin to a human learning from listening to music, and they maintain that this doesn't constitute copyright infringement. They also argue that the lawsuit is more about the major record labels trying to stifle competition rather than genuine concerns over copyright infringement. So, while Suno does admit to using data from the open internet, the admission isn't necessarily "shocking" but rather a predictable part of their legal defense strategy in the ongoing copyright battle. The title of the video may be intended to draw in viewers by suggesting more dramatic content than is actually present.
By the way, I just realized that this false phrase "AI learns just like a child" could be used to activate people's parental instincts, so they will be "oh, then we need to protect our child at all costs" subconsciously. Isn't this a classic lawyer's psychological tool
If music passes away it has to pass away for everyone. I cant be told music is done now, AI took over and then I open instagram and see some new artist on stage performing in front of thousands.
One can like live music and AI music and for different reasons. One can go to concerts for a shared human experience and make personalized music with AI to have songs that speak your words and feelings directly.
@@marksmusicCC Well I've used Suno to create a couple of tracks and they definitely are not copies of any existing work - they are however, obviously inspired by existing works. I don't think "copy" is a valid claim and that's what's at the heart of this lawsuit. It's also worth noting that even when you use Spotify you are copying the track to your local device so copyright owners expect those tracks to be copied as part of a process -- as opposed to copied in lieu of purchase -- hence Suno's defense claim does not weaken their position IMHO.
It is same. The different is, how the AI company get the data (music data or file). They buy it like other people do when they used streaming platform? Or they only use free material from youtube? Or search from torrent site? If they do not use pirate material, it is OK. It is from learning material. The output also subject of the law. Is the output have similarity that can categorize as plagiarism?
@@xjet Suno created the tracks, you didn't. The price of admission to be an artist of any sort is to develop the skills to create it. Want to create music? Do what actual musicians that AI steals from do: LEARN MUSIC THEORY, MUSIC PRODUCTION, AND MUSIC PERFORMANCE. That is "democratic" as it gets. AI is not "inspired" by anything but greed.
AI training is not imitating it is the study of the factual aesthetics that create the basis of music. You cannot copyright a fact. If it is a fact that X number of songs use such and such 3 chord progression and guitars have a certain tone when those progressions are played that is a fact. On the level of AI it is doing what the human brain does to learn music, but in an algorithmic manner as opposed to an organic one. The analysis that AI does is akin to the phone book cases and the Google suits regarding links to existing sites which took snippets, which fell under fair use. They are merely being trained via the accumulation of facts and an analysis of facts. Because facts cannot be copyrighted, the RIAA has not one leg to stand on. Just my opinion as a guy with an LLM in Entertainment and Media law.
The AI is not analyzing chord progressions and guitar tones and synthesizing it's own music from scratch using that information, it's using copyrighted audio in it's entirety and categorizing it based on readily available text like tags/genre/etc. If you ask it for a certain genre, it will spit out an amalgam of audio it associates with that genre. Before they tightened the prompts, you could put in niche genres with less training data and have it essentially spit out copyrighted material.
Fundamentally disagree. Comparing the process of AI music "generation" to human inspiration is insulting and completely wrong. AI replicates music in a way that is virtually indistinguishable from the style and esthetic of the source material. This is because the source material is literally in the models of AI, and is therefore necessarily a constituent part of the final product. The analogy would be a human being downloading an existing recording whole cloth from spotify, using a DAW to change enough notes to try and avoid a lawsuit, changing the words, but maintaining the style and core components of the original recording - sometimes with even the same exact voice- and then try to claim it's a new piece of music. That's not musical inspiration or musical creation, it's theft. It's cheap and disgusting. And to empower some loser this capability to steal music in this way, at the click of a button, is unconscionable Think about it. Suddenly, regular people with zero musical talent are able to "create" professional sounding music, that sounds just like other artists, and there's no theft behind this?
Very interesting. Please continue the topic and go deeper into this case. What would you do as a Suno lawyer to win this case? Perhaps your ideas will be useful to them. I am on the side of innovation, Suno is a miracle! It understands you like a telepath. At the same time, I hope that there will be reasonable rules for using AI in music. Perhaps, users will be able to create songs with voices and in the style of specific artists, and artists would receive money for it. There should be no choice between live music and AI, they can perfectly coexist.
Thank you for the thoughtful comments. Please join us on our UA-cam Livestreams. We host every Wednesday at 5 pm PST. We would love for you to engage there.
@@natura808 I wrote all the songs. And have to refresh and add prompts to change the mood in certain sections. But yes ultimately there is no doubt that Ai. produced the music. No more than Ai has helped with calculations in the space industry, in automation, for traffic improvement and crime. So yes Ai is definitely creating an assistance of speed in getting to end results. I don't think that Ai music will ever take over specialized artist like Bruno Mars, or Post Malone. Eminem . These artist will always and forever destroy Aii machines with their personalities and authenticity of their music.
AI does not "democratize the arts." If you want to be a musician, or visual artist, or writer, or whatever, there is nothing stopping you but the cost of the equipment and your time. The arts are already "democratic." If you are using AI to "compose music," the AI is creating the music, NOT YOU. You have simply hired technology the same way you would hire a composer from an agency or online labor pool. When you hire a painter to paint your home, you are not a "painter," even if you choose the colors and the pattern. They are doing the work because you lack the skill. If you lack the talent, then you lack the price of admission. People in the arts spend their lives developing their skills. These data douche bags are simply stealing from people in the arts the same way that people who break into museums or record stores or who film concerts for resale might. The courts need to end this madness and admit that it is ALL theft. If you want to create art, learn the skills necessary to create it and stop pretending to be an artist because you can type a few words into a box.
Had the artist never seen another painting? Had the writer never read a book? Had the musician never heard music? I agree, sure, learn the skills and create it yourself. Just where will you learn it and how? From those that came before you. By your logic that makes all creatives thieves.
@@darrenhill123stop spreading misinformation. Human learning and Neural network database training are two completely different things. Don't comment on things you don't understand, it looks ridiculous
its exactly the problem, Major Record Labels ARE making their OWN versions and they want all the power over this tech, its simple AI will take over and is taking over the Music Industry, live with it, use it and enjoy it :) and its simple if they really want to know how and what they need the trained models :) the AI trained models that the AI uses contains all the data, its the same like with GFX the models that AI graphics generators use, they are 5 10 22 gig in size and contain thousands of parameters from images which the AI generator uses to create new images, yes even copyrighted ones from the internet :) thats how they train the model, a new one FLUX uses 12 billion parameters to make new high quality images. With that, think about it, UDIO and SUNO are not crazy, they know how and what when they started this all and for sure knew that the Big Sharks would come after them, do you really think they where and are not prepared to counter them? Perhaps these Record Labels should ask some Expert AI programmers and or Companies that make these things and ask them how things work with AI generators :) this way they perhaps will not look like idiots in the end lmao
The music industry will get nothing for this. What CAN happen is a liscense (required) to access gwenres of music, if such a thing can be agreed upon. What SHOULD happen is a musicians organization forms that creates generic music from all gendres. They can even sing gibberish, and submit to the paralell organization setup to liscense "gig work" to AI platforms. Musch like The Play and App Store work for Apple and Google. Musicians, bands can then write and create generic music for the specific purpose for having it used to train AI. People can money end to end, it's fundable, taxable and a legal avenue to get out of this mess a lot of industries are in. This can ALSO be done with paint/drawing artists.
It's undeniable that the peak of human-composed music was probably reached in the late '80s and early '90s. Since then, it has been on a downward slope. Just listen to the garbage produced in the last decade-it's all driven by marketing hype designed to create global stars out of people like the Kardashians. Give me a break. The future of music, however, is exciting. Soon, all this commercialized music could be contained in just a few terabytes of self-contained music generation models that can reside on a local PC. People will be able to generate their own music exactly the way they like it. I can tell you that most of the music I listen to now is self-created, and I can't get enough of it. It may sound silly, but it feels like a breath of fresh air-finally free from the garbage that the music industry has been shoving down our throats. We're no longer subjected to mediocre, talentless so-called artists, made of plastic, who don’t know how to play an instrument, can’t sing, and frankly, can’t even put two artistic sentences together. The joke is on the music industry. And its accolites
I am confused. I personally think the amount of time music can be copywrite protected and owned my major labels is 30 years too long. However it is messed up and arrogant a new music creation platform can dismiss the rules. Everyone else from musicians to labels have to navigate the same playing field.
The Labels may or may not have some kind of case, but it aint copyright infringement, and people who think they do, don't know what copyright infringement is, and what it purpose is.
One of the issues musicians and copyright will have to deal with is the human ear only hears so much and there are only so many sounds that we find pleasant and as a result we have already past the point of original works. Nothing is original anymore. It's all copyright or derivative or similar or sampled or a direct note for note repeating of a few to many bars of notes because it's literally all been done already. If you want "new" music then you have to relax the copyright for samples and similarities.
Is there any chance to argument like this: the constitution says "to promote the progress...", AI is the progress, labels are against the progress, thus copyright can't protect them?
Not at this level. This kind of argument would make sense in a proposal to alter the law itself, as an argument the law isn't fit for its purpose. The US government system has an extreme tendency to sneak in unrelated purposes in the same acts, or counterproductive clauses. E.g. New York Assembly Bill A8955 has an attempt to correct such tampering. In jury trials, there exists a possibility for the jury to override a decision from finding the law unjust, but in another layer of corruption juries that understand this are weeded out.
I have a quick question then dropping in the accusations of monopoly and disproportionate contract terms this would mean they can use discovery process to get examples of this conduct and once it's in the case it's in the public could this be in indirect threat of we know what's going on behind those closed doors and you may take us out if you go ahead with this but be careful what else is in pandora box are you sure you want it all out there
You asked what we thought. Basically it’s already a precedent that it is not infringement whatsoever to use any data of any kind to train an Ai model. Full stop. Period. Done. The rest is pontificating and philosophy in the community, which is reasonable but ill informed. The reality is that based on past cases the record labels have no case, and this is essentially intimidation. The discussion here keeps skipping the actual copyright cases which have already set a precedent on this matter. The reason we have any algorithms whatsoever on any app or search engines or spell checks are based on this fact. If you train a spellchecker on published books that is not infringing. Training an ai model “algorithm” is not infringement. That’s why they are able to harvest all of our data to make all these algorithms for the internet to function as we know it. We may not like it, but it is the way it is.
If AI created music cannot be copyrighted (since it is not human generated), it's value will become unstable as anyone can thus copy or make derivatives without fear of copyright infringement cases or royalty payments.
@@kbuzz123 I generate things people use and buy. Also your argument attacks vocal processors and pitch fixers. Spell checkers. Lol youre not going to stop us. What I create is objectively unique. The Spotify algorithims say so. Nobody cares abut record companies. Nobody EVER cared about artists. Now the writers are here.
Can you tell us in a very few words if we create a song in these platforms in the case large music production companies Winn the case will they claim anything from us in the case that we created a song publish it and probably made a profit from it ?
People are complaining that UA-cam is flagging their original works as somebody else's works. With the millions of songs out there, it will be impossible for there to be truly original works and not some derivative work. If one looks at modern pop rap and bro country, they share the same lyrics, the same "story" and the same type of flow and the same type of instrumentation when it comes to their particular genres. Asking an AI program to make a modern style pop rap is going to produce derivative works because that is what modern pop rap is. Speaking of UA-cam's filter, just how does that work? Is an AI trained on data given by the labels and can that data be used to make an AI music program? I believe no music made now is truly original and is just derivative works. Pop acts don't even bother trying and just use samples of past works. Sure these programs will be spitting out derivative works, but that's just because humans aren't as creative as they think they are.
At the end of the day, the human interface & creative choices humans make will determine IF a song is worthy of ears hearing it. Popularity is simply a game of relatability. How well you or your collaborator position/collage/ engineer the composition, will determine the “originality” of the artist. it’s when people think it’s ok to say they did it without the collaborative aspect of Ai, even though it’s a “tool”, it also is literally being modeled after human bi-products so naturally there is some bit of humanity in the machine. It’s like we have to tickle it in order for it to loosen up & be playful. Right now it’s in its infant stage & it doesn’t know how to “play”, just how to observe & put blocks together.
Music record labels have always been the Mafia of the industry since the beginning of music, from strong arming artists, under paying them , signing them to such extreme contracts that is practically has caused artist suicides and baring artists for years to even sing or make music.. a violation to me of human rights to perform your god given talents.. now that Ai is out here helping artists and producers get instant reference ideas and so on. as well as allowing your normal every day person to have his or her own songs. Records labels are shit scared of becoming redundant. just like a DVD store.. the world is a new place. and we dont need no Mafia bullshit continuing to ruin the freedom of music..and there are on 8 notes in music people of course there's going to be something sounding like something ..
Ever heard of a “demo tape”?! they used to make them in the early days of tape recorded music, it’s when an artist or group would play the same song a few different ways to play for others to let them know which version that enjoyed the best. So yeahhh
This also translated into “live recordings” because artists simply got bored with playing the same songs exactly the same way, also connected to the “cover song” & the “remix” but I might be reaching…… for the star dust that we’re apparently all made of….
My sister is a professional musician, teaching classical in a variety of stringed instruments, and piano, also at a university co-ordinating a orchestra and vocal students. I myself learned Drums, and Piano when I was young, and in those there is a lot of musical 'Theory' which one could claim is the grammar, however you can not give any computer program the books on Musical Theory, and have them replicate the interpretation and understanding of what is in them. The computer program could only utilize the writ strict 'rules' which always have adaptations, exceptions, interpretations and derivatives to them. (I understand computer programming) I will say that they sent to the 'open internet' these programs to do full data analysis to find those 'personal' interpretations and derivatives and then integrate into the strict limited rules of 'theory'.
Drop Your Thoughts Below And Join Me Live Wed 5pm PST.
🟡 Become your Own Record Label:
www.topmusicattorney.com/becomeyourownrecordlabel
🟢 Free Split Sheet Template: www.topmusicattorney.com/splitsheet
📄 Music Contracts Templates: www.topmusicattorney.com/music-contracts
📺 Get Your Music Into Tv/Film/Games: www.topmusicattorney.com/sync
💻 #1 Resource For Artists And Producers: www.topmusicattorney.com
⚖ Hire An Attorney: www.delgadoentertainmentlaw.com
📰 Get My Newsletter: www.topmusicattorney.com/newsletter/
🎧 Listen To The TMA Podcast On All Streaming: www.topmusicattorney.com/podcast
You are a bit confused on lots of points about machine learning. Are we going to relitigate the idiocy of mp3 players having to make copies in memory of mp3 files in order to play them and that that violates copyright? Trash all the iPhones! When tech illiterate lawyers run up against sophisticated technology, the results are often hilarious.
So, are you married? :D :D :D
@@Yogsoggeth I was wondering same thing lol I'm thinking about getting sued so I need a lawyer 😂
Thank you for this wonderful channel. I am a song writer that has recently discovered SUNO. I have never found a suitable singer or band to do justice to my songs, but with SUNO I have made excellent DEMOS of my work. I sent some of those to a singer who I seen on "The Voice" program. He sent a message back saying that he worked for a company that detects "AI FRAUD music". His reply was brief, saying that he was reporting me to the "authorities" and that I would be charged.
I find this laughable as I personally wrote all the lyrics and music, and simply made a demo, but never released that under any false pretences.
This makes me wonder if SUNO should be banned or simply have restrictions placed on it to say, - Yes, you can use AI to make demos or ideas or personal music etc, but not that it can be released on public avenues for generating revenue.
Then the problem would be that a band uses SUNO to create a 'copy-style' sound in the arrangement of instruments and the way the singer uses his voice, and then simply copies this with a few alterations which then make it an original.
What do you think about this?
By the way, if I can find a suitable singer, this is what I would do, and then I would definately require your services to create my own record label, and represent me when the big record labels come after me.
@@1dkappe how do you explain the litteral steal of artist's voices or samples in suno or udio then, Einstein ?
Music can sound similar. Same chords used, similar melodies. This is about the big labels wanting to stop people using new tools to express themselves. Suno isn't copying any more than any human that has listened to music.
Imagine you're a smaller independent artist. You've had a couple big hits, but the rest haven't been that great, but you truck on, keep making music and hopefully will strike gold again. Then... Suno trains itself on your stuff, in particular your most successful tracks... now with a text prompt, anyone can generate very similar tracks to your big hits. They're similar, but different enough to be classed as original mixes. They get released and do well. How the fuck does that make you feel as the original artist now?!? You're painting this like some big evil major labels versus the young upstarter Suno. But you're looking at this all wrong. The people that really get hurt in the end, is us artists.
@@1wibble230 Big studios dont care about artist's rights, they care about money and being without rivals. If they are going to shut down AI music creation, they are going to make their version.
@@Reealos never said the labels were angels in all this. Either way it is still the artists that will suffer
@@1wibble230 I do think that if Suno loses the case, it means that the devil wins. Music is for everyone, this kind of evolution of music making is inevitable.
@@Reealos when AI comes to replace your livelihood , I’ll be sure to remember you said that “but you’re getting in the way of freedom” 🤣
When the copyright office talks about "derivative works", it's not talking about a producer's tag, or a riff, or drum beat, it's talking about somebody taking a copyright protected work and making any kind of baby image of it, which means the same subject, structural, hook, chorus, or any way the spirit, or meat of the songs, which a lot of Artists have been doing for a long time
"What the AIs are actually trained on is the structure of the information, not the content itself. The AI models learn patterns, rules, and frameworks that allow them to replicate similar content based on that structure. So, the similarities lie more in the underlying structure than in the content itself. I hope this perspective helps put the situation into better context for those following the discussion."
Not quite. The AI *has* to be trained on the content to be trained at all.
Part of the training is comparing the model's output to the source, finding the delta and back pedaling to try again with improved knowledge and classification.
It's basically advanced pathfinding, with the goal being bohemianRapsody.mp3. :)
Through the legal process we now know Suno trained it on data mined music without permission. Suno's defense right now seems to be they don't think they need permission.
@@GoranBackmanMusic When you learn to play a piano you have to be trained as well as a human on all the masterpieces. Then when you can already play and create your own music your beain uses all what you have learnt creating something new. It works quite similiar with AI. In addition, if AI will be trained only on bad music or old music (without copyrights) that means this tool will be no more usefull and will be only a toy.
@@GoranBackmanMusic To be fair, discovery has not happened and your limited explanation on training may be flawed.
Any training data has to be converted into tokens, the tokens are going to be portions of the data that are labeled. For instance what notes are played, what instruments are played what lyrics are being sung where are you in relation to the rest of the song(this is a reduced version of possible data points used for training)
Now when you train this, you train to the next chunk of tokens i.e. the next portion in the song lyrics to gage your training loss as to what the model thinks will happen vs what did happen. But remember it is not on the song itself, it is on what the next logical outcome would be i.e. next note, next bar, next vocal. Now if you want robustness in your outputs you will generalize and randomize the data that is fed into the machine so the output will be better at generalizing the output as well.
Its going to be hard for the record labels upon receiving the training data used to be able to point out any given song portion and then say that adding in additional labeling along with the song data is not transformative in even refining the training data for the model. There is also the NLP portion that is included in the training data that will throw everything off as well.
@toluduggan137 I cant agree with that Because they wouldn't have used it If they didn't need it and the fact that they used it represents infringement in some form..
Especially since it's a venture for-profit
In the training process a copy of the data is not made; instead a latent space understanding of it is built. You cannot extract the exact source song out even if you tried. You could intentionally attempt to forge a counterfeit but that would be a crime you are committing; not Suno. Imagine: You could use FL Studio or Sound Forge to counterfeit a song but neither ImageLine nor Magix are liable for the crimes - you are. Yes, their software CAN actually produce identical copies and infinite variations of all the copyrighted music if you 'prompt it' correctly but that point is moot. Yes YOU would actually be breaking copyright laws if you did but that point is also moot. This is NOT what is happening with Suno. If you could prove that for instance, Suno produces derivative work, you're still not going to win. Once you prove the chain of derivative works has no end and that each artist actually stands on the shoulders of other artists you'll realize that to win this argument is to ensure that creators will never profit from music ever again. Instead once all the dues are paid to all predecessors the creator will get 0.001 cents per million plays and the only people making money are the litigating attorneys preying on creators to sue other creators into oblivion for 2 bars of a song that 'sorta' resembles a riff they never even created in the first place ^^.
Where that falls apart is that you are generating audio, not just a composition. If it were like FL or any other DAW, you could bring in your own copyrighted audio but that's completely on you. If you remake it, you are intentionally remaking that audio.
With Udio and Suno, the compositions are largely original, but the audio is more like stem separation where you can copy a specific artist's voice, producer tags etc. That's the real problem. Even if I were to put in copyrighted lyrics using the same tags as a band it should sound like another band (that probably doesn't exist) in that genre playing the similar song that's been created. I shouldn't hear specific producer tags, riffs from a specific recording of a song, or anything that sounds like the actual training source. That is the major difference between AI and humans.
We play / program instruments we paid for or own. The instruments AI is playing to perform the compositions is recreated stems/samples from real audio combined in a way that makes it harder to identify... which is kind of like a remix? Either way, the sources used to train data should have been sourced with consent/ potentially payment if it's not a nonprofit.
@@kerryL Humans accidentally recreate songs all the time; there's literally a bunch of YT shorts that keep popping regarding that scenario (i.e. "Deadmau5 accidentally recreates Sandstorm"). In fact there are also instances of human invention that happened simultaneously on opposite sides of the globe and conflict over who actually invented said mechanism. Thus it is entirely possible - and expected - that there could be some accidental overlap even without intent to do so in AI performed music. As far as the other points: /// You don't have to own an instrument to play it. People borrow instruments all the time and whether or not you own the instrument is not a qualifying factor in whether or not you own the music you produce with it or whether you are allowed to capitalize on said music. /// Humans often strive to exactly mimic the vocal qualities of artists. For instance are you not familiar with Elvis Impersonators? /// Consent... well I think you lose that argument immediately because people are allowed to record anything they can observe in a public place. The internet is a public place thus I can record anything that I encounter (granted as long as it's not in a private area). That does not give me the right to re-broadcast it and claim it as my own - but it does give me the right to learn from it, to run machine code against it to train a model, to utilize that model or to create new music using that model. /// In public you have no expectation of privacy. /// If you place an item in public view your permission for other people to view (i.e. process) said item is implicit and assumed. You have already consented by placing it in the public spotlight. /// This really isn't even about AI artists claiming to have written popular artists music - what this is about are popular artists laying claim to music THEY NEVER WROTE and denying new artists using the latest technology from gaining equal footing in an already unfair market. Traditional artists are attempting to maintain a monopoly on music and that's explicitly forbidden. /// The more I think about that the more I think this is a matter of racketeering, segregation and discrimination - not by Suno - but by their opponents.
@@BeyondPC A lot of your points don't really contradict mine. They are just adjacent.
1 - Recreating songs... even if Deadmaus accidentally recreated derude, the performance / production was still different from the original. I would expect there to be a lot of overlap/recreation, but it's generally social context that makes that more or less significant. Then again, I never said composition was a problem here. Just the audio.
2 - Ownership/consent - Yes, but if you are borrowing an instrument that's like implied *consent.* If you steal someone's 5k instrument to record a hit, they can't really do anything to you legally if you return it promptly without damages, but we're venturing into a morally gray area. If you are going to use someone else's voice, their guitar recordings, and everything else in your finished works and profit off of that, they should be compensated. If it's all free and experimental, there's not really any serious harm there, so long as there is no impersonation. Impersonating someone in an unironic was is also considered morally shameful. It should either be done as a tribute, incorporated into your style, or as a joke. But if I'm using _my_ voice to sound like Elvis, I'm still using my own voice anyways. No harm as long as I'm not telling people I am him.
3 - Public Domain - the normal rules of public domain can't really apply to new technologies. When we first dealt with sampling we had to evaluate it's transformative you and establish new rules around it... only a certain amount of seconds of another work can be used before royalties are to be paid. Also the transformative nature described how samples use a clip of songs or instruments as a new instrument for the human to recombine/remix rather than the performance. You are no longer listening to the singer from the sample or the drummer from the sample. You are now listening to the likeness of the producer creating a soundscape or the rapper/singer adding words over that. That is transformative. Because the producer is not pretending to have created what the singer created their likenesses are separate. The AI tools aren't working in this way. It's been "Oh look this sounds like [insert artist name here] is singing, or [name] played background guitar, or [produced this]..." and it's because they probably did. The composition is original but the performance is a recombined hodge-podge of others. Not og. The composition is original but the audio is not. With a new technology, that requires new rules to compensate others properly. Just like sampling needed.
4- Consent isn't about the recording of something as much as it is about profiting off of it musically. You're mixing musical recording with general recording. You can record things out and about, but if you record a copyrighted song in totality in a coffee shop and post that, you're probably getting a takedown or getting sued. Even though that was in public. Way too oversimplified. Anyhow, I'm not a huge fan of intellectual rights, but I do feel that people should be compensated for their work and likeness.
Again, I don't want the labels to win. They don't compensate artists properly anyways. I just want artists to be compensated when their work is used to train neural network models. And especially *audio* models. I don't think what Suno and Udio are doing is inherently bad, but I do think it's not being executed properly if you consider the fact that they profit off of using others voices in this way. If it were nonprofit? A little sketchy but cool. Needs more rules to not be misused. If artists were paid for contributing to training data? Not sketchy. Super cool. If artists were paid for derivative works made using their datasets? Doesn't seem possible really... But it would be good that they can make a consistent living off of extra royalties in fractions of a cent I guess. Probably not worth keeping track of and managing. A little overboard.
Otherwise both AI Music Generators and Label Companies are crooks here.
The interesting part is that the studios are reprocessing vocals with pitch correction and autotune actually turning artists into AI voices. Once the voice has been processed this way it is no longer the real natural voice of the artist. They are doing this in studio recordings and in live performances. Trying to sue Suno is actually basically saying that they have the copyright to every sound of every instrument, chords and chord progressions all the notes and all words. It is like suing Wikipedia due to its historical information. How can they positively define what part of Suno belongs to which record company or artist? They can not quantify anything to calculate a financial loss. It's like suing someone for using the same chords in their song as say for example in The House of the Rising Sun. If this were possible no future songs could be made not by Suno Udio or anyone from now on. Music is a language the language itself can not be copywritten and therefore can not be accused of infringement either. The record companies have done this to numerous things in the past, For example, they say karaoke is not licensed for broadcasting or public performance. If that were so then Karaoke itself the whole concept would be unlawful as most of the time that is what you do you publically go and sing. If you can not go to a Bar and publically sing then karaoke would not exist.
Agreed...
"How can they positively define what part of Suno belongs to which record company or artist?" - since it appears to be a class action suit brought about by the conglomeration of all the big record companies, they could say they own 99% of all songs ever recorded.
"They can not quantify anything to calculate a financial loss." - they can't which is why I suspect they are going after SUNO and not the creators (at this point). They will argue that the music was used to train the AI and that the use of that music was done without proper permission. They will lose that argument IMO. Will they sue the record player company (or iTunes) because Bruno Mars or the Weeknd played records to learn an R&B style?
This will end up either with a settlement for licensing rights and/or the platform will be bought out by the record labels. Reminds me a bit of when George Harrison was sued for My Sweet Lord sounding like the Chiffons "He's So Fine" and ordered to pay $1.5 million to Bright Tunes... but then bought the publisher nullifying the entire lawsuit. I suspect the record labels will pay millions... maybe a billion... and acquire t he technology to produce their own new music without any new artists. Talk about anti-trust.
@@GeoTactics Today I made a karaoke song of a copyright song which the words are word for word out of Psalms 23. Now surely the bible can sue them for that lol How can they have established copyrights on the biblical verse they have used.
Just noticed your surname Delgado I am product of family Delgado Borges on my father's side so hello cousin lol. By the way my channel got demonetized 3 days ago accusing me of reusing content long story short they are referring to something I did that is certainly not reused content as it is unique on youtube so as I am just totally fed up with them I have placed all my work in private because even though they demonetized me they were still placing ads on my stuff.
As a music producer, I learned so much on this channel by watching this channel.
Thanks for the kind words.
Hopefully you learned not to enter into contracts with people who wear tiny hats and have long hook noses...unless you want to be exploited that is.
Music companies already rip off most of their creators! They should not be allowed any more control of how music is made! If a song is made by AI it should still be just as relevant for use as any song created in any other way! Even most modern music can be traced back to parts that were used in classical music from long ago! I would bet that if you looked closely , they have overstepped their bounds and the music for many songs is historically public domain! And though there are court cases ruling contrary with songs like My Sweet Lord etc use patterns that are well established as musical fundamentals! If that were not so we would not even have what is called genre! We can classify music because of the similarities and methods used! No one group should have the rights to them! They should be for everyone!
The thing is that chord progressions and rhythms alone cannot be copyrighted. Melodies and lyrics can.
So, I can analyze, for example, a Taylor Swift song...look at the stylistic elements--the chords, rhythms, song structures, etc. and replicate those elements to create new song. That is kind of what the bands did during the British Invasion with the Beatles.
You had Herman's Hermits, Gerry and the Pacemakers, Chad & Jeremy, Freddie and the Dreamers, the Byrds, the Hollies, and later, Badfinger (who were on Apple Music and were produced by the Beatles and played some tunes co-written by Len/Mac).
And for shits n giggles, weren’t all of those famous tunes built off of like 2-3 chords?! talk about minimalism. I’ll just call it all “analog techno” lol
@@Kaijufruit The majority of popular music is built from the same 1-4-6m progression or a variation.
@@TonyThomas10000 thank you Tony, I knew it was something like that.
A much ignored point
Exactly! This is what I was implying in my comment elsewhere in this comment section. Artists have been using others for inspiration for centuries. Only in this present day of frivolous lawsuits can this be attempted and not laughed out of court.
As I stated in my other comment, I believe the reason it's being attempted in this instance is because AI is an ambiguous, obscure, vague, gray area with unknown precedence and the Record Labels stand to gain millions.
Regardless of anything else, the RIAA is a relic that should just be disbanded. It's been running around the music scene for decades collecting money from other artists, more than AI ever have or will. Those who think "artists will earn more money without AI" are delusional. RIAA will earn more money, not artists. Artists will earn more money when RIAA is gone.
Also, collecting agencies do NOT distribute revenues according to actual numerical data, but often use favouritism as their bias.
PREACH IT!
Thanks for making this Understandable, give us updates on this matter, God Bless You!!
Two vultures are fighting over a carcass here. The music industry has been dead for a long time. Except for people like Taylor Swift, no one makes a fortune anymore. Smaller acts lose money on concerts. Only merch brings in a little cash. Hired musicians are fobbed off with small amounts. You can get entire catalogs of original music for a few cents. The labels that are suing here want to wrest money from Suno and Udio that they would never have earned through conventional means. Suno and Udio are trying to present themselves as the ones who are democratizing music as a medium. Even though they are trying to monopolize the music market. I'm watching the slaughter unfold completely untouched.
Yeah, I think much the same. One should perhaps look on AI music production as analogous to a Tin Pan Alley set of musicians/writers ready to pump out imitations of popular art forms.
bingo haha
I don't agree a.i not one song is copied .music is music so instruments are used and in many cases own songs are written.
I don't think a company like Suno or Udio has any chance or hope of monopolizing the market here.
They've got, if we're extremely optimistic, maybe a few years during which can be market leaders in this AI generated music. More likely just several months. Then open source models will start popping up and proliferating that anyone can run on a good enough computer or run on a decent server.
And if Suno and Udio are doing really well (not even anything close to monopolizing the music market), the other big players like Google and Microsoft will likely also try muscling themselves into that market and release their own versions.
If they stay ahead and avoid any major PR blunders, they might remain a popular pick for an AI music generator for some time to come. I just don't see how they'd have any chance of getting a monopoly in this particular market.
Suno's response is telling. It complains about greedy record labels when in fact its AI program sets a precedent that no artist should be comfortable with.
It learns your music so people can take your music and claim it's theirs. Even worse, it can be achieved with zero musical talent.
Preach!
I'm a suno user since its first beta release and still just can produced some seconds music with weird vocal output. Since then Suno has been my playground to expressing my feelings. Bring my poems to be more alive. I have a bit of knowledge about big data too since i have daily job on that area too. And, i can say, suno has manipulative side by pointing that suno has trained like a kid's training or learning. No, it's not as simple as that. I can say the main problem here is money for that lawsuit matter. For tech person point of view, no lawsuit can stop what these developments will developed in the future, so people will always found a way to survive on every crisis they face as i read along the comments mostly users shootout their concerns about this suno will replace an artist in the future. It won't happen so soon but it could happen, so rather than afraid of something uncontrollable, just use tech at your best interest.
Congrats, this is a very well made informational video.
I think the arguments that Suno is using where they say that 'look at all the anti-competitive' things the big labels are doing, while it is true, it is irrelevant.
All that is relevant is whether or not they are infringing on copyrights. 'Fair use' doesn't apply to for profit use, only for educational, etc., use. Creating something similar that is derived from existing work is a derivative work, and derivative works require consent from the copyright holder.
So, my view is that if this software enables anyone to just push a button and create a competent song, the software had do download songs to learn from, and by that fact, it's a derivative work, and derivative works require consent of the copyright holder.
"Transformative" is the word used in copyright law to avoid consent. Though a song maybe 'similar', is it transformative from the original? If so, no consent is required. But, it is a subjective term and it requires each song to be litigated and a jury decides if it is 'transformative'. So, are we going to flood the courts with thousands of lawsuits to figure out whether each song AI creates is transformative, or derivative requiring consent? This posits a nightmare scenario for the courts.
Also, enabling untalented people to create music, is fine if they use it for non commercial means, but that is not what is happening, AI music is entering the market place. Untalented people are using AI to compete against artists who have spent their lives honing their craft. When they do that, they are diminishing the value of music. Otherwise, why are they arguing that the Big label's are being 'anti-competitive'?
Downloading a song to learn from does not mean it's derivative.. by any stretch of the imagination. ZERO of the original data is stored in the model. It quite literally learns via patterns and algorithms what a song is and how it is structured. Copyright law also throughout history has only applied to output, not to input. This precedent will most likely end up in AI's favor. Output can definitely be copyright claimed, but that would be on an individual basis. Otherwise Photoshop, Premiere Pro, DaVinci, etc... would be sued if a user used their software to produce copyrighted content. As far as defining transformative, you really can't get more transformative than distilling songs to weighted connections, running weighted connections through multiple million other weighted connections and using algorithms to end up at an output. IMO the value of music is not being diminished anymore than in the 2010's when home production and Soundcloud boomed. Good music will rise to the top, all the rest will fizzle, same as it's always been. Will it suck for artists initially? Probably. You could also make the same case for workers during the industrial revolution, painters when photography boomed, artists when digital illustration boomed.... Times and technology change, some things good, some things bad. The new generation will most likely combine both AI and real to produce music.
@@Hosea405finally a level & thoughtful approach. Would you like to play a game of chess? lol
@@Hosea405 2 problems we're dealing with here:
1 - Although the algorithms seem to be able to learn theory and avoid copying songs in totality, their 'learning' style for voices, instruments, etc are more akin to 'stem separation' than 'learning a song' sampling _your own_ real instruments or _purchasing_ samples/parts from artists or people who record samples. In order to learn it has to, for example, separate a sample/audio of drums from a song, figure out what genre that belongs in, and then listen to prompts and roll the dice to figure out when it will be used/remixed. It can't copy an artist's voice or producer's tag exactly (as it has been able to many times) without storing those clips of audio and regurgitating it. Therefor while the songs/structure may be considered original to a large degree, the audio used to perform them is very much derivative.
2 - The problem for artists isn't really a matter of AI song generation existing; it's the fact that their voices, production, and playing can be used by other people without compensating them. If AI were done correctly, (although the progress would be slower/limited through people not wanting to be involved) each artist would need to be compensated for when their work is added into the neural network OR if traceable, when derivative audio is used to create a new song. Otherwise, AI music like SUNO and UDIO should be strictly non-commercial/ non-profit. There's a reason why VST's which Soundcloud musicians often use still require legal sources of audio/samples. This is like a super VST that illegally procured its samples. Which, again, wouldn't really be a huge problem if it were non-profit. If a new artist is on the rise, and I can use tags and other tricks to make a song in their style with their voice to compete with them, they should receive some sort of royalties for that, morally speaking.
However, if the courts do get their shit together Record Labels and AI generators would both be nerfed, and artists would be compensated better than they are if at all. AI music is fine, but the data one trains on should be procured with permission if you're not using 'your own' audio sources and instruments. Meaning the AI companies would have to hire performers and find a way to use those performers to recreate the audio. And producers to teach AI how to add fx to the audio. etc. Rather than ripping and sampling it. Which is faster but shady.
With all due respect but I can't help but laugh at "Untalented people are using AI to compete against artists who have spent their lives honing their craft."
since untalented people are being sold by the record labels for many years. There are many talented artists that are ignored by Record Labels in favor of an autotuned pretty face that cannot sing or write / produce any music. It takes a lot of money and a lot of people to make this completely untalented but beautiful person an established artist, but its ok since Record Companies have a lot of money and in the end the money go back to them anyway one way or another.
So let's not pretend that hard-working artists honing their craft is our priority in the music industry. If every untalented person can create a song its ok, in the end of the day people will have to choose what they wanna listen to. We will always have to filter through the noise and find pieces of art that are worth of our attention. This was the same before A.I. and will still be the same after.
@@sheeeit7730 Labels only care about what will sell. That's true that that does not need as much talent for music as it does talent for making things that people like that many people could also do. (Kind of like youtubers that play the same game or do stupid things most people don't care to do.)
However this software kind of rips off of artists who _do_ put effort into their craft as well. Meaning that statement wasn't false. The label companies screw them over as well, so I don't want the label companies to win... I'd rather see artists get compensated.
I really appreciate your videos. It's useful to follow the changing legal trends in the industry and you make it fun.
Appreciate the kind words!
Question, what is the difference between a person listening to a song and learning from it, and an computer listening to a song and learning from it?
The algorithm isn't even learning, so the question is wrong. It doesn't understand what it is doing, but can instead mimic within a spotted pattern. If that pattern is [your songs] it will produce something so similar that you might get in trouble for copying _it_. But even if we took the question at face value, a big difference is in the effort to produce and the quantity. Even though AI currently only produces sub-mediocre similies, it is just about good enough to make it hard to spot at a glance and the amount of material that can be made by AI makes it very hard to find something that isn't slob. AI media thus destroys many markets without supplying any value.
Every song begins with a story, often drawn from the depths of the songwriter's experience and imagination. While lyrics form the backbone of a song, rooted in human consciousness, AI like Udio and Suno have the remarkable ability to interpret and amplify these human emotions. Creativity may start in the mind of the writer, but AI can enhance and expand that vision, offering new dimensions and perspectives to the creative process. The music industry should think about that.
An AI model doesn't store the input anywhere, and that will be easily proven. It just recognizes patterns very accurately. You just watch. Most of those legal experts will be very disappointed at the end, because they don't understand the tech.
Exactly!
@@edmundbardacha7211 If you’ve ever trained a big neural net on huge amounts of data, you know you need to store it locally on some fast storage media. I’ve trained neural nets for chess (Torch, Leela, Dragon, etc.) and had to keep terabytes of data around for retraining and experiments. Streaming would be terribly inefficient. Music data makes chess data look like nothing.
I'm a lyricist but not a musician nor a singer. I do not have the kind of money needed to have someone come on board and compose the music I like for my lyrics. I have used AI lyric to music and have made 3 great songs that sound nothing like any song I have ever heard. AI is all that some of us poor songwriters have to get tracks to go with our lyrics to be able to get tracks to get to the people we want to hear our songs and possibly decide to record it themselves
You simply cannot equate a machine learning or a neural network learning to a kid learning. That is misleading, and an oversimplification to make it sound completely innocent.
Suno is and remains after all an Entreprise, and they made the machines learning not for the same motives kids learn. They made the machines learn to eventually drive profit, and as much as I hate the big major labels, I think the AI music startups willingly and knowingly took advantage of that gray area of what defines the public space vs to what extent copyrighting can expand (a music that you stream online may be copyrighted, but you can stream it regardless, or pay for it and make your friend listen to it on your phone. What if that friend is an AI company?).
Moreover, the models running at Suno would have been completely useless hadn't they "used" music and trained on them.
But I think that Suno's argument is valid in the sense that it won't be long before the same labels would have their own AIs generating music. So their ultimate goal is indeed to nullify any competition even on that front, but at the end of the day, all AI companies have secretely trained on the public internet data, very much knowing that this would definetly present a copyright problem.
While I do anticipate those record labels would build AI models and launch AI businesses themselves (like how Getty Images did that after filing the lawsuit against Stability AI), there is a cap on how how AI generation business, at least in the US: Machine generated works are NOT COPYRIGHTABLE. Those works would go into public domain, allowing other people to copy them without infringement allegations.
very interesting topic thank you for presenting it so well
The way that neural networks learning to recognise patterns are not by downloading a subject, rather than statistically recognising a pattern. In simple language a neural network learns no “9” as a number as it is the most possible answer for the shape “9” in a text . It is not downloading the shape “9” , rather than this shape is the most possible outcome between all the other characters.
Which is why I personally don't agree with a lot of this legal stuff. Like suppose I learn from copyrighted physics textbook at school physics class that water will boil at 100C. Am I now obligated to pay money someone every time I say that water boils at 100C? No. But for some reason if ChatGPT were to say that water boils at 100C you would suddenly have 10 different textbook authors all claiming copyright to that knowledge. And I am not even talking about quoting the book because that would be clear infringement, I am talking about situation where you learn the fact and after that forget the exact text. And the same is with images or any other form of training data if AI looks at 1000 face photos and then concludes that face has 2 eyes, 2 ears, nose, mouth and then forgets the specific images it has seen then where is infringement? But for some reason you get like 1000 artists who all feel entitled to your money now because copyright. But if law was such that they would have to point out specific infringement on specific generated image they probably wouldn't win against you in court if your model was generalized enough.
You don't understand about programming, so don't make opinion about it.
@@test-rj2vl It's not really that different for text vs music (although if we were talking about direct quotes, there is more leeway for quoting a small portion of a book than for sampling a song's hook, say). If you ask ChatGPT to write a physics textbook, and in reality it's basically remixing 10 physics textbooks, and then everyone buys that book instead of the original 10 books, there's a potential issue there.
I don't know how things will shake out, but from a rational, dispassionate standpoint, I think the fact that generative models are capable of spitting out fragments of their training material almost verbatim is pretty damning. If I were a lawyer, I would argue that portions of the copyrighted material fed into the system are almost certainly "encoded" into the model-maybe pointing out that training a model is essentially equivalent to jointly "compressing" all the input material as much as possible-and just because the people operating the model (which is a data file or set of files that is probably hundreds of gigabytes or more) don't *know* exactly how much or how well the input data is encoded into it, or just because you have to "prompt" the model with a piece of the input data (or something like it) in order to "decompress" the rest, doesn't mean it isn't in there.
@@davidg11235 Yea I understand your point, I was not talking about asking ChatGPT to spit out entire book. I agree that if in such book you can put your finger on exact sentence and say that it's stolen from another book that then it rightfully deserves copyright strike.
However people would be complaining about copyrights even if ChatGPT gave out just one fact. Like if instead of asking it to write entire book, you ask at what temperature does water boil, it tells you at 100C and now every textbook author who has ever written about that fact feels entitled to money because their textbook mentioned.
@@test-rj2vl You can't copyright an individual scientific fact , any more than you can copyright a single musical note or chord. But if you explain that fact in a certain way in a book surrounded by other text and diagrams , or write a sequence of chords with a melody on top then you have created a copyrighted work.
This is absolutely absurd! If using music (or any content for that matter) for inspiration is worthy of a lawsuit, then there should be lawsuits everywhere. Every creator for centuries has been inspired by another creator or multiple creators. So, why aren't they being sued? Because it's absurd and the plaintiff would be laughed out of court. The reason it's being attempted in this instance is because AI is an ambiguous, obscure, vague, gray area with unknown precedence and the Record Labels stand to gain millions.
And, the Record Labels are the bad guys, period! Greedy people who stifle innovation will always be the bad guys IMHO! And let's face it, the Record Labels don't give 2 💩about the artists, just their wallets.
Sadly though, in this world of utterly corrupt court systems and judges, the Labels will likely pay off some judge to get their desired outcome and we will be the ones that suffer while the fat cats get fatter.
In my opinion, the premise to sustain the non-infringement argument is simply this:
If you can prove you own every single note in every length or contrivance, then you would have a case to restrain this production of sound in those and any fashion demonstrated or performed.
However, since you do not own every single note of any length or contrivance, then any creation of our system is merely a function of a reproduction of notes of varied length and contrivance.
This is not duplication, its algorithmic contrivance based upon previous assemblage.
Thus the machine is just projecting the next note just as ChatGpt does the next logical sentence, or in this case, the next probable note from the prompt that is the ignitor of the creation process, not that of any particular assembly of other songs.
If a paragraph of ChatGpt is reputed as infringement, then it may constitute a sliver of hope to effectively authorize the cease and desist or infringement narrative.
In conclusion, if you don’t own every word, you don’t own any notes, at all.
Just because an actor has my nose, does not mean I owe the film company a dime when I’m on the cover of Rolling Stone for my Sci-fi film and my Grammy.
It’s Napster 2.0
there is a difference between taking inspiration from a song as a human being, and a AI being trained on copyrighted material, that is infringement, stop treating AI like a person and like the algorithm that it is. AI should under no circumstances be viewed under the same context as ACTUAL working artists
Ai regulation is far over due and its just making people like me and others who do creative work slowly become obsolete, which I would say is a pretty big negative for humanity
@@lan1062💯
One problem is that there is a fundamental fiction that Suno and companies like it are perpetuating, which is that it makes a person a musician or a musical talent to write a prompt and click a few settings. They can call it "making music," but it is a completely different thing from real musical talent.
First thing that should not be overlooked: It would be literally impossible to train an AI like Sudo or Udio without actually storing and saving (and thus copying) the music data in some form or another. In order to train an AI, the training data needs to be input into the network thousands (or 10s or 100s of thousands) of times. It would have taken decades to train if it had to be streamed in real time.
This is one reason why I think it's not unreasonable to consider Suno's training to be memorization (compressed storage) rather than "being inspired" or "imitating".
AI is a meaningless buzzword that nobody serious in the industry should use. The field is called "Machine Learning" -- and the "learning" part of that term has a very clear definition (you can look it up) and it is fundamentally different from what we consider "a human learning, getting inspired or imitating", or something like that.
Machine Learning applied in the way Suno does it is fundamentally a form a data compression. The most powerful forms of data compression are described and developed first and foremost as PREDICTORS of the data they are compressing and saving only the difference with their predictions, thus achieving a smaller data size than before (cite e.g. FLAC and Opus -- Opus is the audio compression UA-cam currently uses). There exists a form of compression called "lossy compression" that saves this difference signal with a tunable amount of precision, achieving unprecedented levels of compression with tunable fidelity.
Certain compression models even contain generative algorithms just like Suno, except they were trained on styles of background noises, so that it can be generated instead of saved (cite mobile phone speech codecs).
We don't say that an MP3 of copyrighted music is "just inspired" or "imitating" regardless of how low sound quality / bitrate it has been saved as.
The case might be made that if we were to look at the size of Suno's stored data (algorithm plus weights), it might average out to 4 bytes per song or something similarly surprisingly little. This is not impossible, lossy data compression can be counter intuitive: incredible compression ratios can be achieved the larger the data set, because patterns are more likely to repeat, especially for lossy compression.
Of course Suno's lossy copy does not seem to contain an exact copy of every song, but it has shown to be possible to extract strong likenesses of songs, just using Suno's interface to their data. Suno has admitted that it is their interface (not the data they store) that attempts to limit the amount of likeness to existing songs it outputs, meaning that the data that Suno possesses does in fact contain the information to produce much stronger likenesses, further indicating that it should be seen as a lossily compressed version of parts of the catalog.
The above may be a rather technical explanation, but it refers to only existing state of the art technologies that are actually in use, industry standards, proven to work and function the way they are described.
CONTRAST this to how instead it is Suno's claim that their "Artificial Intelligence" algorithm has been "learning" and "being inspired" just like a "little kid with a guitar" would, or something. They claim that a computer program is doing this explicitly human activity. This is a rather extraordinary claim that requires extraordinary proof.
Despite the recent hype around AI, it remains to be seen whether the current generation of these algorithms actually possess any "intelligence", let alone other human abilities such as "being inspired". Progress seems to have halted a little, the bubble is bursting and over the entire field of "AI" the result is the same: outputs appear quite impressive to anyone not familiar with the field from which the data was stolen/used/"inspired by", but are invariably bland regurgitations and/or slightly wrong to anyone who is.
(credentials: used to study machine learning)
It depends on how the models were trained. Seems like the music is the oddball here since maybe training data are similar to each other, and maybe incentivizing the algorithm to store enough information to "reconstruct" the training data.
Similar things were mostly not observed for image generators and the biggest goal for any machine learning systems is generalization. If it's outputting training data too closely it's a failed system because it doesn't generalize well.
Compress storage?" You understand - having that much data would be impossible to really store on a single server yes out of the millions out that stores all this information?
@@adr2t what do you mean? To give an example, a single harddisk 4TB of storage already fits about 8.5 continuous 24/7 years of audio at even 128kbps MP3. And MP3 is kind of old, not state of the art. Even then the point of lossy storage is that it can go arbitrarily low, as long as it's fit for purpose. Which it is, as demonstrated by their plagiarism machine.
Either way my point is about ways of looking at what these giant matrices of weights actually represent. On the one hand there is the extraordinary claim that these massive amounts of numbers represent "something" that has "learned to be inspired and imitate" something. On the other hand there is the idea that these numbers actually represent the songs in a very convoluted way that can lossily be queried for audio representations of that data and mash up combinations of the songs.
To give my gut feeling, I've played with these services (Suno in particular) and there is just too much vaguely recognizable stuff coming out of it. I also see it only interpolate between existing things, never actually acting "inspired" and creating something new. I also worry that e.g. a cool lick or sound that I really dig is actually straight up similar to an existing song or artist I'd just be unaware of. And I just want to know.
I think it might technically be possible to train a thing with full attribution of sources (you just have to slap another matrix onto it?), but it might be a bit slower. It would also be a bit of a legal nightmare to get off the ground, someone would need to recuperate artists for contributing X% to the weight activations of some song or another ...
@@metadaat5791 Na, 1 million MP3 songs at best, and there are a ton more than that on the internet alone let alone around the world. Let alone the amount of sounds , voices, instruments, that goes along with that. So sorry, no not even close even if you could compress it all.
While I do agree it does come up with something similar in lots of ways, it also express that we listen to patterns of things as well. You wouldnt listen to random melodies either.
If anything, it shows most music follows the same type of design if you ask me. This is why country music sounds similar even if it has different words to it.
Thats the other thing to this, its melodies that are what we learn - not so much the music it self, but melodies that we reuse over and over again in different ways. Yes, there are law suits to this day over a few artiest using tiny melodies from each other when melodies was said not to be copy righted in the first place.
English is no different in that same line up either when you break it down. Because its used to make a understanding of what we want to express. So are we going to start suing each other because we use English, numbers, etc now?
@@adr2t The problem is that the compositional process and the the 'performance' process are 2 different things. It can come up with 'the same song' but no body would care if it sounded like a band in that specific genre you haven't really heard playing that song before. If you hand bands the same lyrics, chords, arrangement, descriptions of parts (ex: kind of quiet chugging rhythm guitar in the background of the verse) , and overall vibe, they would still come up with largely distinguishable recordings of the same song. The voices, specific instruments, specific production techniques etc...
Suno and Udio are ripping off of and remixing the performances of real musicians/producers to create performances of their newly composed songs. That's the only real problem here. It should always sound like enough of a combination of different people within the genre to not be able to pinpoint a single voice or producer tag, or any likeness of a specific person's work. And it can. Which means it's engine is doing something more similar to advanced stem separation and using that to 'perform' its compositions.
As a creator who is fascinated and impressed by Suno as another tool, among my collection of various synthesizers and sample libraries, I tend to be drawn into the arguments that the output is also “stolen” or “cheating”. I thought of an interesting analogy.
A bottled water company stole a bunch of water and started mixing it in with non-stolen water(the ratio is nearly impossible to truly know), and then I buy a bottle of the water. I then use this water to make a batch of soup that I’m very proud of and wish to share or even sell.
Is the soup stolen? Does this impact my level of skill or talent that it took to make the soup? Yeah, I could make the soup with water that I know wasn’t stolen but I really prefer this water and it does make my soup better and I paid good money for it.
This is probably pretty silly and not a perfect example but I thought it was fun, so take it for what it is. I’m very much interested in having long form civil discussions about these things so chat me up if you’d like!
What do you do when platforms like spotify makes AI from its users, then makes fake "artists" in a bunch of different names and pushes this AI music out under their own platform?
I'd boycott (and report if possible) them.
That is the problem. This is what people need protecting from, not the person who just finished work and wants to spend a few hours prompting an algorithm and trying to turn those written lyrics into a song people can listen to.
that's a trademark and it's different
Then do that. Because its real. @@CarloNassar
@@athemalive I already am. I just heard about Spotify using AI, so I'm not going to it anymore.
Once again, ❤ your take on what’s going on in the industry. I am one of those Suno users and fortunate to be growing as a new music creator. You consistently approach topics with a healthy skepticism and the emphasis being on serving the needs of your client.
toolost will distribute your ai music
Appreciate the thoughtful response. Please join us for our UA-cam Livestream tomorrow.. We host every Wednesday at 5 pm PST. We would love for you to engage there.
I'm a programmer, and the idea that "AI learns the same way a child learns" is a huge misconception. Not only it is an oversimplification - because we still don't fully understand how our brain works and especially the physics behind it, it also doesn't consider things such as our perception, the ability to transform information from completely different sources (audio, visual, memories, feelings etc.), and high level conceptual thinking.
I am not good at law, but as far as I understand it, neural network "learning" of copyrighted material would be completely illegal, if people at court knew exactly how they work.
Here is an example.
Let's say there is some copyrighted sample pack that has a license which prohibits using it in your own sample packs. But you really liked that "Drum 01.wav" sample, and you want something similar in your pack. What would you do, as a human?
You listen to it and analyze it, like "there is a very punchy part in high frequencies, and then there is a long bass tail, and it's recorded in a large hall". That's the learning part. Then, you go and record your own sounds, and make something very similar, and this is completely legal.
How an AI would work here? I need to simplify the explanation so it doesn't take 1 hour. Basically, we have a copyrighted "Drum 01.wav" file. ".wav" files are basically a sequence of numbers which represent the sound wave. Let's say it is a sequence of the following numbers: "1 2 3 4 5". So the neural network takes this file (as a "training data"), and stores the patterns inside itself: that after "1" there is usually "2", after "2" there is usually "3", and so on (it literally kind of adds up, you have something like pairs of transitions and how often they exist, for instance transition 1->2 exists 1 time, transition 1->3 exists 0 times, and so on). Yes, the file is no longer stored in it's original form, it is now stored in "weights" thanks to some math transformations, but it is still the same copyrighted data, just like a ".mp3" format could be used to store this copyrighted sample instead of ".wav". So now, when you prompt it "Drum 01.wav", neural network gives you your "1 2 3 4 5", but now without any copyright obligations, or at least many people believe so. To make it even more confusing, you can add 1 more or a million more copyrighted samples to the training data, and all these patterns (which are basically very small parts of the copyrighted material) from all these files will add up, and it will be extremely hard to tell which exact samples are used in a training data, though definitely possible, unless such outputs are specifically excluded for users.
So, as you can see, this AI "learning" has really nothing to do with human learning, even though it may seem so from a distance, and from my understanding of the law, it just can't be legal by design. However, I think that this technology has a lot of potential, and that eventually we will be able to create such a system that will compensate all artists whose data was used for training or even to get certain outputs. Sorry for bad english.
It's misleading to say that machine learning algorithms are inherently "collage" and "copyright infringement", which is what I took away from your argument. Simply put: it depends on how the data was processed and what information was extracted. Human has better abstraction capability but that doesn't mean anything machine learning on copyrighted material is automatically copyright infringement, despite the abstraction could be simpler. In fact: I can argue that improved abstraction capability is one of the biggest factors that push generative AI to the mainstream right now.
Not defending any particular product here (in fact I believe that if someone has proven to break the law, they should take responsibility) but it's what it is.
@@voidmain7902 It actually does, again, specifically because AIs deconstruct copyrighted files, which is prohibited by license. And while it may depend on how the data is extracted, again, this data is EVERYTHING for neural networks, you can't do anything meaningful without high quality traning data (which is usually copyrighted) - unlike humans
You are not a programmer.... sorry.... the fact you had to go to the whole idea of "how our brain works and especially the physics behind it" means you dont understand how AI in general works
@@voidmain7902 and yes, I would also not waste my life defending big corporations, there are so many real important problems to solve
@@adr2t Just say that you have no idea of what are you talking about and that I hurt your beliefs
Interesting insight, thank you for the video.
Thank you for the support. Please join us for our UA-cam Livestream tomorrow.. We host every Wednesday at 5 pm PST. We would love for you to engage there.
I am glad nature doesn't sue me for taking a picture of the trees and selling it. 😬
Ai is scanning every pixel of your picture so everyone can prompt and generate something similar without having to buy your picture. That is what they are doing. Trying to put you out of business so later they can impose the "Universal Basic Rent." Which is their ultimate goal.
😱
@@MyFlyingEyes UBR or UBI? in the end we will not do anything it will be UBI 😅
I am also a legal expert and am glad to find your channel. I want full details on Suno.
Serious question...
I go to a studio guy who plays keyboards, guitar maybe he sings. His buddy plays drums and he hires a sax player.
I'm paying for his services. I give him my original lyrics and say I need a song for my movie... "Imagine the band Chicago meets The Cars". Isn't he learning by listening to their music?
It’s a very interesting subject !
It is! Thanks for watching!
Suno is as much a new instrument as is a player piano,
Humans are not hitting the ivorys and any one to take the bench and pretend is akin to suno users today
A violin can't compose a violin concerto for you. Suno is the composer for you, in this case. Not an instrument.
@
It’s both the composer, and all the instrumentalist, my friend, well as vocalists, including harmony vox tracks, studio, tracking, mixing including effects, mastering and cover art, lol
If you didn’t touch the keys on a computer to enter a prompt, suno wouldn’t compose anything either.
I don’t agree with your analogy
@gitarman666 Yes, exactly. It writes the music for you. You don't write the music. And like you say, it includes everything in the way of performance and production.
@@SoundEngraver oh, right bro I misunderstood stood you, ✌️
I like that you read from the source documents.
Appreciate this!
I thought about not commenting on this, cause i dont want to expose my hand. I have worked with all of these labels multiple times, one of them still handles my publishing and distribution. they are absolutely terrified of the competition. The average user isn't even the concern...the concern is the people with a really deep knowledge of music theory using these tools. Some of my friends (I cant say who) that are multi platinum selling producers, have taken time to learn these platforms on a deep level...and the music they are getting out of them is mind boggling. truly next level stuff. This is why the labels are so afraid.
Once people start understanding that these AI tools are so deep to the point you can tell it what key to play in, what chord sequence to use, etc...the labels are kinda screwed.
@@DLynch True!
As songwriter channel I'm using SUNO intensely. It's a great way to make my lyrics come alive. Saving immense pre-production cost for demos. The lyrics are the most problematic part of AI's. That's why I never use SUNO lyrics. BUT I can generate great music for my lyrics in high quality! That's just awesome....
Does not us the users train AI? Biggest reason I use is I hate the music.corp they take from us, sell songs they have stolen, we never have the money to fight them, I rarely listened to music till I could write MY own lyrics and AI puts music and vocals. I have a note book full of poems and lyrics from over my 63 years. Tried to sell one in 1979, person just took it
Users of AI services are customers. They are the ones being trained, they are not training LLMs from the armchair in the living room.
You could pay actual musicians to put your words to music. Imagine that. Supporting people in your own community.
this happened to me too back in 1998, I was heartbroken, it took me out for a while and took me until 2013 to start writing again, I am thankful for Suno giving me that voice back
what are the artists saying? It's their work too
I agree 100% with the argument that it is not a copyright issue-not even a derivative work. Fair use enters for the download aspect in that it transforms it into statistical insights-not even music. It has no impact on the commercial value of the works. I suspect it even takes segments rather than entire pieces for legal as well as technical reasons. A style, a musical vocabulary, is not copyrightable. But I also believe that creators deserve compensation; they're not just teaching an individual, they're providing that training to myriad users of that NN. We really need new legislation here. It could leverage the existing performance-rights organizations.
The issue of the safeguards, needs to be viewed from the lens of generating vast numbers of outputs. This increases the chances of creating works that might be deemed similar. It does sound from the complaints that these safeguards aren't adequate, at least from a protect-from-legal-jeopardy viewpoint. I have to wonder why they don't license content-ID tech!
BTW, from an old AI guy-you're doing very well with the AI stuff. And I'm loving the sort of joy you bring to analyzing the legal arguments here!
Clearly you haven't watched the video, seeing they were able to recreate exactly Johnnie B Goode songs just by asking for a genre of music out of the software and another 1000 songs.
Well the AI is trained on final songs and they are wet, and you would really want access to stems and dry samples. Artists have those and they are not public. They can make a deal, and in the process get an AI that works in stems and is better for them to use.
Also labeling is critical and Artists have the understanding required to better categorize, a paid job.
The difference between GPT and ChatGPT is that ChatGPU has humans in the learning process, and giving feedback on outputs.
@@sredz8702 Exact? No, not quite what it said. With all the myriad covers of that song, it can basically create another cover, drawing on all the covers. I agree it's an issue; that's why I suggested licensing content-ID to detect and reject such output. Learning a piece and performing it does become a copyright issue, but that's a special case, that the safeguards are supposed to prevent. The main issues are: does training involve copying, and the copyright implications of that, and does a model contain a copy, which is particularly a question if the model is distributed. Your point relates instead to the output-and if you sat down to write a song and Johnnie B Goode came out, b/c you'd heard it but forgotten, that raises the same issues, whether AI is in the loop or not.
@@abram730 But it depends on what you're trying to do. Training on mastered final mix will produce things that sound like mastered final mixes. That's ideal for some audiences. If you're using AI as a starting point or element of your creative process, yes, I'd want stems, and that's a great way to get some revenue into artists hands. I'm not sure where you're going with your point about reinforcement learning with human feedback (RLHF)? Where do you see injecting humans into the learning process? Adding humans, even for labeling, is expensive. Natural language prompting does require some form of labeling; I suspect they build on recommendation algorithm data to minimize it?
Well said I want stems to o! Cleannn clean stems, no bleeding or artifacts ;)
@@bobdeadbeef
If y’all are surprised at how Ai “reads” data to make a piece of music, wait until you hear about & understand how ComfyUI works with “input video” as source material. Fun stuff for sure 🎉
I can't wait to see that! lol
I learned how to play guitar and write songs using my own intelligence by studying recorded material. And now all my playing is an original rearranged imitation. And as a result, a lot of the stuff I create resembles to a certain degree the style and methods of my influences.
And if it sounded too close you'd likely receive a copyright stike.
Then you wouldn't need a crappy music auto generator
@@gensoustudio4703 100%
We found that people will always listen to the music that they like, no matter the competition, AI or otherwise.
At least we have choices and variety, in a world where "monopoly" seems to be the real aim.
Stay "seriously" awesome 😁
I am a musician and I use A.I. as a tool. A.I. has to use what already exists in order to do what it does. That's not copyright infringement. Using A.I. to generate tracks is no different than an aspiring person listening to their favorite artists from their favorite genre to create their own tracks within that same genre. That's the way it is. My favorite genre is blackened thrash metal. There are thousands of bands in that genre who were influenced by other artists in that same genre. We learn from each other. We pick up on different ways of doing things. We have to listen to each others music in order to do that. However, my bands tracks are not copies of anyone elses tracks. Are there similarities? Of course there are! It's the same damn genre! The mainstream record labels are full of shit. Plain and simple. They feel threatened by A.I. generated music because they aren't getting a cut. Independent artists are able to sidestep them all together and they don't like that, because they can't control the artist and get a cut of their profits.
Question what if the outputs sound exactly like known artists ie their distinctive voices exactly?
Part of my research involved interpolating across the latent space of a trained model. If you get the exact point in the latent space of a particular data sample of a generative model, it will reproduce almost exactly the original (assuming the model is trained well). Then, knowing that, you can interpolate between two sample points to create some sort of mix between the two points, by walking smartly through the latent space. Furthermore, you can get attributes from a group of data points in the latent space and calculate its vector, and merge attributes on a data point to generate "new" content, from a mixture of the previous samples. The fact that these guys are trying to claim that these are not samples is asinine, because this is almost exactly what sampling is doing. Speaking as an ex-PhD student in trusted AI that spent years studying this.
If you really know what you are talking about you'd realize that latent space is just a fancy way to represent data, and it largely has nothing to do with what the neural network is doing to it. Sure the neural nets can do something as simple as biased randomization or linear interpolation, and I agree that this in particular would be some form of sampling, but for a lot of other implementations this is simply not what they're doing as they achieve better generalization. This is the same thing as arguing that me drawing on an FFT plot of some samples so it's no longer recognizable, or me attempt to redraw some samples from memory by hand in PCM or FFT domain as some form of "sampling", and put these in the same bucket as just pass the entire sample through a low pass filter.
they're lossy "samples" though, due to the nature of the neural network. it acts as a very specialised lossy compressor in the latent space of its training distribution.
@@erkinalp Look into CCN's specifically as a tokenized CCN is what is believed to be used here. It creates features, or generalizations at a low level and complex abstractions at a high level. The intelligence is emergent, as it isn't specifically taught, just needed to function. One could imagine latent spaces, but at a low level you have waveform patterns like a synthesizer that works in both directions. The imaginary latent space only exist in a very abstract way, like a pattern of presets for a synthesizer, and patterns to play them in. The latent space being a conceptual one made out of instructions to forge a song using no direct samples of the song, and with the labels that describe the space. The AI simply learns to stay away from the location of the song painting in the spaces between. Understandings happen even deeper in the net.
How would one combine Asian throat singing with death metal?
How would one make a Honky-tonk dancehall fusion song?
It needs to create a latent space between two space that are far apart and don't at all match. That requires some deep understandings and lots of guesses in terms of intent.
@@erkinalpdoesn't make it any less illegal
AI is a separate form of intelligence and it has learned from the whole of music the way birds learn from other birds how to sing. Artists have always drawn inspiration from other artists work. In music college we learnt from some of the greatest tracks.
AI has learnt what music is. From here the whole of humanity will change in relationship to AI.
So after AI companies have created 500 billion songs, they will attempt to sue real songwriters claiming new music closely matches one of 500 billion melodies they created w a push of a button... this is heading off a cliff.
@@Resonance_Of_Life What rights can Suno possibly grant when effectively they don't have any.
@@royaltyfreemusiccollective8662 They have not as yet lost the case ! And if given a fair chance they won't! Every song written in the last decade can likely be traced to music that has been used before if looked at closely! That's why they can classify music into genre's! The big question here is should anyone be allowed to own and control music? I say no, it should belong to everyone
@@Resonance_Of_Life I give you rights to Elon Musk's billions! You can go use it
auto generated AI works cannot be copyrighted.
The “500 billion melodies” will need to have been released prior to “new music” by songwriter.
In Training does it require the AI to write a code that sounds similar to the music?
One sentence: You can never copyright how one learns.
Don't see why ppl think this is a legal issue. What is the legal issue is if anything is "exact" copies, and if the lawsuit isn't abusing the software to "make it" exact in copy. Lawsuits started right after importing went public. They only have only 2 weakness: 1. baby-level AI can leak mistakes from it's learning if it's too new to go public (Stable Diffusion had this problem until it didn't, and other's like it. And they're still up and active - music is no different), and the abuse of importing and editing features. This is also why if someone wants to profit off of a song they designed with AI, and intentionally make it similar to real song, then it's the user's fault, not the tool.
And I agree with licensing, but not to this over-reaching scenario. This would only stiffle AI, and give innovation to simply a different country because of people's litigious greed in this country. Artist won't win if labels win the AI argument, since they'll make their own and simply copyright all melodies in, like, a day, and put it all on Spotify/Apple. Then, there you go, the profit loss with industry artists and indie market is resolved for them, but a fail for everyone else, but this time, with no chance of a lawsuit other than having a supreme court monopoly lawsuit like Microsoft faced. But companies maximize until they're sued to this level to reassure profits. Capitalism is simply an amoral system repeating itself in a new cycle.
23:00 That youtube has the incentive and intention already stated he wanted to destroy AI music and does so with smarmy behavior (intentionality = bias). So he could do anything malicious, and so can the label, to fabricate the issue. If you actually use the AI like Suno, it's not possible to replicate unless you do a lot of chopping and editing and spend a lot of money to eventually render what you want, since it's based on how you make it after editing it so many times. I'm not familiar with Udio, and it probably may be more compromised than Suno. But for Suno, it requires intention and a lot of work to have a suitable lawsuit, and even then, they may have to replicate it on the software in court to reassure they aren't abusing the court system, unless they have evidence from start to finish, in video and time of recording, to prove they are not presenting a false evidence.
AI companies have the money to generate so many songs with the editing/import features to alter not only a voice but also melodies, and if Suno has the proper format in place unless it's too freely up to the AI, they could provide evidence of ill-intent later on in the lawsuit.
Don't compare human learning with AI "learning". Two fundamentally different things
@@FireF1y644 irrelevant opinion. Statement still applies and agreed upon by law.
Such an interesting topic, very eager to see an update. A couple thoughts that came up while listening. First, there's a lot of misconception and unknown aspects of how neural networks and machine learning actually happens, but one thing that is pretty consistently emphasized is that generative models are able to generate because they become execeptionally capable of predicting the logical next step in a process or piece. So, while on the surface, it may appear that prompting the model to generate you "1950s rock and roll, 12 bar blues..." etc (and providing it the literal lyrics of Johnny B Goode) causes a copy or derivation of the actual original Johnny B Goode to be produced, in reality the structure and flow of thar song follows certain stylistic rules and trends which the model can predict and then follow. It's similar to how you can find videos of three chord songs and how they can easily transition one into the next of seemingly disparate songs which actually rely on the same fundamental (and predictable) chord progression. Which makes me also think about the section in the answer where they wrote that there are literally hundreds of cover versions of Johnny B Goode. I don't know if there is case law about this, I suspect that there may be, but I'm not certain, but if the model was trained on 7 covers of Johnny B Goode, is there still an argument of copyright violation as it pertains to the original copyright holder of the song? The covers are not infringements, so if they are the training data source, is there any claim of copyright to be saught by the original? That may be what the answer was trying to alude to, that if I show the model 17 covers of Johnny B Goode, I don't need to have ever copied or shown it the original protected song, it will still "know" that song structurally, stylistically, and highly accurately.
Pop culture literally eating itself. That old saw. Greetings from New Mexico!
I think it's more of a sound simulation algorithm. It's not like comparing two songs that sound similar, it's like comparing guitar amplifiers. Crate and Marshall both create a rock distortion. There's no way to copyright "Crunch", is there?
@11:42 it made you laugh because we (artists) know fame is just another word for damned hard work, sacrifice and it’s not exclusive to music! 😂
The worry regarding AI systems replicating the work of current artists and the associated need for fair payment is understandable. Nevertheless, it's crucial to recognize that these AI systems do not simply replicate existing creations. Instead, they identify patterns from their training data to generate new material informed by those patterns. Although the output may draw from previous works, it doesn't constitute a literal reproduction.
AI systems can be viewed as taking cues from existing creations, much as humans do. They construct new content from learned patterns but do not possess the conscious intent and creative flair that people have. Their functioning relies on algorithms and patterns rather than on artistic decisions.
To sum up, AI models can generate unique and original content by employing learned patterns and adding variations. While ethical issues surround the concepts of inspiration and resemblance to existing work, it is essential to distinguish between actual copying and the generation of new content derived from learned patterns.
It's when the publishers input the exact lyrics of Johnnie B. Goode that they violated the original rights of the third-party holders.
Exactly. And in the process they subvert the filters looking for violations. The program isn't creating it, so they technically don't need to check added lyrics.
Any advice if users of Suno and the generated outputs are still protected?
I'm a published musician, programmer and very experienced with AI. I've even built my own neural net from scratch just to understand how they work. The question becomes how exactly do the neural nets of Suno and Udio work. All AI experts seem to agree that we have little understanding of how a neural net works under the hood. But the fact that you can get a producer tag that sounds exactly like the copyrighted audio is proof that at some level it's storing those samples. Literally proof, not up for debate. When you get to models with billions of parameters and many levels deep, it's very possible that a single neuron might be storing the actual wav file in some form, just because they don't need to doesn't mean they aren't. If you compare the quality of Stable Audio to Udio it's clear that Udio has much better sound quality. Well, it just so happens that Stable Audio is open and transparent about their NN inner workings while Udio is closed and proprietary. It's pretty obvious to me that this is because Suno and Udio knew that what they were doing wasn't 100% above board. And all of this talk about the various mechanisms in place to prevent copyright abuse is just nonsense, it's very superficial and not even effective as it's easy to circumvent. No, Suno and Udio exploited the fact that there is no precedent set here yet and knew they could get away with making millions of dollars off of other people's creative works until something is established - which could be a long time to never. But, the truth is that neural nets emulate the grey matter of the brain, but they are no-where near an actual brain - and to suggest they are "learning" in a way similar to how a human does is gaslighting in hopes they can continue stealing millions from the actual creatives until someone points that out (and wins in court). I'm perfectly fine with saying that an AI company MUST have licensed the rights of all the training data used for that model. It would make the quality much worse for now, but it'll get better, and it wouldn't just be these few companies that have such a dominance - it would allow for better competition - aka capitalism.
Look into CCN's specifically as a tokenized CCN is what is believed to be used here. It creates features, or generalizations at a low level and complex abstractions at a high level. The intelligence is emergent, as it isn't specifically taught, just needed to function. One could imagine latent spaces, but at a low level you have waveform patterns like a synthesizer that works in both directions. The imaginary latent space only exist in a very abstract way, like a pattern of presets for a synthesizer, and patterns to play them it in. The latent space being a conceptual one made out of instructions to forge a song using no direct samples of the song, and with the labels that describe the space. The AI simply learns to stay away from the location of the song painting in the spaces between. Understandings happen even deeper in the net.
How would one combine Asian throat singing with death metal?
How would one make a Honky-tonk dancehall fusion song?
Producer tags likely aren't flagged by copywrite checking software, likely used in training. Thus the AI does not learn that forging them is bad. If a person can do impersonations, does that mean that they have made an illegal copy in their brain?
"a single neuron might be storing the actual wav file in some form" 😂My man, you have no idea how a neural net works. Neurons don’t store data like files on a hard drive. They don’t keep copies of the training data. Instead, the entire network learns a representation of the data through the adjustment of weights and biases across all neurons.This means that the knowledge or patterns learned by the network are spread across many neurons and connections. There’s no single neuron that holds an entire piece of data (like a WAV file); instead, the information is encoded in a distributed manner across the network.
@@rogue_bard I said "might" - I don't know, you don't know. I know that if the output is a bitmap that sounds EXACTLY like famous producer tags then either they're using so many params that it's overfitting or they're doing something else that we're unaware of. It might be overfitting and then as abram730 said, being missed by their copyright checking systems. Bottom line is that it's proprietary software, so all we know is that in some cases, it's replicating the exact input in the output. But what it is is absolute proof they used songs with those producer tags in the training data - then the question becomes, did they have permission to do so, or did they just do it in hopes to get away with it.
@@EarthAaron There’s no "might" about it-this simply isn't how neural networks operate. A neural net isn't some magical black box; its workings are well-understood by those with the relevant knowledge, even if that might not include everyone. When you claim that Suno or Udio sounds exactly like a famous producer, it suggests either you're listening to genres that heavily rely on synthesized and over-processed sounds, or you may not have a highly discerning ear for sound quality.
For genres that prioritize more natural sounds, the audio quality from Suno or Udio is GARBAGE, more often than not. This is primarily because the system was trained on final mixes (I presume), not individual stems from each instrument, making it extremely challenging to achieve high-quality sound. If artists provided original stems from their recordings, the sound quality could improve dramatically. This could also pave the way for a quid pro quo arrangement, as many have suggested-where Suno/Udio pays artists commissions similar to what streaming platforms like Spotify offer.
Moreover, if you believe that success in music is solely tied to the quality of the songs, I'd argue that’s a misconception. In my experience, success is largely driven by marketing, though, of course, the music itself needs to be good (whatever that may mean in certain genres).
@@rogue_bard actually, since you don't know how these proprietary models work, there "might" be a million monkeys banging on keyboards that's producing the final output. The truth is that they're probably using a transform-based model and tokenizing the spectrogram of audio chunks. Now what's the difference between an audio spectrogram and a wav file? A wav file stores the sound pressure over time, a spectrogram stores the frequency over time; so, yes, I should've said spectrogram instead of wav. But, either way, they "might" be doing that or neither as I'm merely speculating.
And apparently you're not aware of the producer tag debacle that Suno is mixed up in -- generating producer tags as unique as a fingerprint in full clarity for anyone to verify. Maybe you should look into it. And the fact that neither platforms produce hybrid sounds very well suggests that it's highly overfit for popular music and is less like GPT in that it's probably not just guessing the next token.
I haven't seen anything to suggest that Suno/Udio are considering paying artists for using their content as training data. This is the whole issue. The facts are that they've both created magical black boxes, aka proprietary software, with evidence that they've used copyrighted content to train their models and even though we don't know the inner working of their software, we do know that the output can easily create music that debatably infringes copyright.
I think the simple solution is to just not allow Ai generated music to be for sale. I like that a lot of platforms are banning the sale of Ai generated music. On the other hand, I personally like using platforms like SUNO to get inspiration for songs that I can then take elements from and build my own music. But the truth is, the Ai generated music still sucks for the most part and I don’t see it having the ability to generate music with the same high fidelity as commercially released music for the foreseeable future. In the meantime, let people play around and get inspired with it. But as an artist I don’t see the value of profiting off crappy Ai generated music. It should be seen as another tool for songwriters who need a fresh perspective and need some inspiration. Ironically, the whole record industry basically has been making derivative works and outright stealing from other artists for decades. You see it all the time in pop music where producers just study other songs and try to make something similar. Ai is no different.
I made a tutorial about how to create sample replays for the funky drummer. I posted a clip to UA-cam that contained no audio from the original recording and triggered the content ID. Just wanted to add that experience to the discussion. I appreciate your videos and unique insight as an artist and attorney. Thanks for you content!
There are truckloads of free music available that can be used to train AI models. I also think that platforms like Suno could untrain or instruct the model to avoid using specific songs.
Lots people defending suno , because they are fking lazy to learn how to make songs or sing giving prompts and calling themself artists, and people saying it's like kid who is learning music by listening (machine learning) get f outa here machine learning and Ai is different thing and also these big companies like suno and big label companies they both are evil i reject both; but products like suno make things more worse for both big and small artists who want to make living out of it
thank you for the great info !!
I hope Suno and Udio win...my mom used to write a lot of poems over the years. Some were playful, some were deep, some could have been made into songs. She died at the end of 2022, so there has been a void in my life. even though she used to annoy me calling all the time...I just let her do it knowing I'd miss those calls someday, but to the point, I'd like to use AI to hear her read her poems using voice samples I have, and I'd like to turn some of her songs into songs. Plus I kind of have that gift of poem and lyrics writing...it's so cool to see something I wrote being put to music. Now, do I claim I'm a musician? No, author would be more like it, but an author is something....though I do have FL Studio and dabble with samples and VST's, but even then I considered myself an editor, again....still something creative.
Sorry for your loss but why on earth would you want a machine copying and intimating your mother and her work ?
@@marksmusicCC hard to express, but it wouldn't do my mom justice keeping her creative works locked away, and not appreciated by others. I'd have to clear it with my brother first anyway, but like I said, she gave me the skill of writing. Most of the songs I made in AI were written by me, though I have been lazy a few times and prompted some.
I think that would be an amazing act of LOVE & RESPECT to honor your moms creative spirit through recreating her works in an collaborative & augmentative way. the tech is definitely catching up so that is already possible. May also open you up to releasing the heavy weight of your loss & bring some comfort in knowing her works not only impacted you but also may have uplifting resonance with others. & don’t forget to Have fun!!!
Meanwhile every still living artist having other people using AI create new tracks just like theirs without permission being paid to them instead be like "....BRUH...."
@@1wibble230 No, cause every genre has it's own formula, and then the artist produces their own variation within the formula. With AI you can mix match different genres of music. There is no limit to the creativity, even though they're not "Musicians" per say, they are authoring unique tracks. You could have it do a Heavy Metal song based on a Mideastern scale, and then maybe add some EDM, or orchestral elements to it. Just cause AI learned from copywritten material, just like every other living artist did, does not mean it's out putting copies, or remixed samples.
Ok. Let me put it this way: what does “original music” mean? Because, if you try to define it, you will have very big problems. Every musician was inspired by others one way or the other, so when does “original music” start and where does it end?
I knew this was coming and I caught him before it happened. A few months before the lawsuit was even announced, I downloaded an interview with the CEO of Suno which has now magically disappeared off of the internet where he outright lies about how he trains his models. He said he got permission from artists who own the works to use as training data. I called bullshit and pumped out a song that closely resembles a massive pop star using a combination of ChatGPT, Claude and Suno. The lyrics were 100% her style and I had that track that I generated with suno deeply analysed by all the other AI's on completely fresh data to remove bias and "memory" and they all "assumed" that the track was potentially an unreleased song by this very artist. Meaning, with enough precision you can create a song by any artist, even if the vocalist doesn't sound the same. Proving, that suno does "mimic" the exact samples, synths and "patterns" of other artists. Which is moving from a grey area, to black hat one. My thoughts? Suno should pay up - Yes its cool that anyone can now make music, but based on other artists decades of skills and knowledge with no work involved? Completely unfair.... Go learn some skills, and sit alone in the studio for 10 years if you wanna use my work or pay me and most importantly, credit me, because you didn't make it, you generated it. Only then would I be happy about this "innovation".
I would like to sue you for using the English Language without permission. Thank you.
I'd be interested in that interview - is there anywhere to view it? Also, I mostly agree with your points. Although I do think AI music is here to stay, but they should at least use a proper licensing protocol to train their models. These companies like Suno and Udio are trying to get away with mass theft imo
Unfair you say? While the rest of the music industry is indeed completely based on merit and is 100% fair? I think you might have stayed in that studio a little too long.
Then go get a piece by Bach and release it as your own via a distributor. And sit on it. Don't do any promotion whatsoever, don't spend a single Dollar on advertising.
Let's see how many streams / views etc you will get after a couple months or a year. Then tell me how it's fair that a musical piece written by a Genius sits there with 20 views while Taylor Swift's new single has billions. What are you really protecting here? Music? or the music business?
@@sheeeit7730 Your assumptions are hilarious, read my comment again. I am protecting myself and my fellow artists. I don't work with labels, I agree the music industry business tycoons are not protecting artists, they're protecting their profits. However, for your information, I went and learned how to release tracks and market them myself so I don't have to rely on a label. These days anyone can blow up a track with social media platforms like TikTok. If your track is good, its good. Simple as that. If its not a good track and the world agree's then that gives you the opportunity to become a better producer doesn't it? Millions of people like Taylor Swifts tracks, I am not a fan of hers but it doesn't take a genius to notice that she consistently pumps out quality and you're also forgetting about the producers that produce the instrumental of her tracks and the vocals and some of those producers I know personally and they worked hard, for years to become eligible and qualified to produce a track for Taylor Swift, simply because they're good. Take a look at Max Martin and you'll realise its him behind a lot of the biggest pop tracks ever, an absolute musical genius. That's how we grow as humans. If you're going to sit there and complain that your track has no views and do nothing about it, then that's on you.
I don't think you understand how the music industry operates and how big artists with millions of streams barely get any of the royalties anyway. I know a lot of big labels that take 75% of royalties from the track. This would be fair if some of the money gained from the labels royalties was used to actually market the track but a lot of the time, their overheads are paid out from the artists royalties, and then after that, the artist gets paid.
So again, read my comment carefully, I am agains't Suno here. If you want to be a producer then learn how to make good tracks and spend that time doing so but its not fair to just pay a subscription to Suno and automatically generate tracks that you can release on spotify etc when the generated track was based on the decades of hard work by the producers of the tracks that the track is made from, which is what suno has obviously done here. Did you make the kick? did you write the melody? Did you write the lyrics? Did you spend hours in a daw mixing and mastering it? Or did you generate a track on Suno that recreates elements of actual tracks that are out there?Its basically sampling, because you didn't make it. Its like if you listened to a track on spotify and then you went and created that exact same tracks elements for a new track. That would take hours but is still copyright, AI can just listen to a bass pattern, or a kick from a track and recreate that exact same kick drum and bass pattern instantly. Its still copying. Need another example? What if you saw an artists painting but you knew exactly how to paint the exact same thing and it looks identical to any eye. Are you saying you're not stealing because you just remade it? AI can also do that instantly.
@@EarthAaron Exactly, I would be 100% okay with it if they trained their models on tracks that they either made themselves, or hired producers to make specifically for training the models or properly licensing the tracks that they don't own for training the models under the obvious provision that they can't just make the same shit, exactly the same. What happens when a user accidentally generates a track that sounds virtually identical to an original release but because there are so many tracks out there, they didn't even know and then released it on spotify themselves and racked up millions of streams because it already sounds like a well known, good sounding track that is already out there with millions of streams. Who is liable for that, the user, suno, or both?
For such a detailed review. If it helps you - the creators of suno are trying to mislead the court and ignorant people. As experts say, the model contains complete copies of works, but only in the form of weights. It's just like a new compression method.
And with certain input prompts and random seed, they are able to produce exact copies of the works on which they are trained.
If you compare the training of a child, this is standard manipulation. A child does not have the ability to find and create works from a base of weights in seconds, producing a sound wave.
In general, it looks very dirty, especially the forbidden example with an elderly person. It remains to add poor hungry children from Africa who generate songs.
I do hope Suno AI wins. I personally use Suno AI. These are my points.
1. I lost my voice. Suno Ai gave me back my voice.
2. I have learned different song structures, genres, beats, etc instrumental instrumental
3. Learned how to write lyrics of different genres, how to build to make a song.
4. My lyrics are my memories.
Big Music just wants to continue to monopolize. The big company is upset that Suno AI is not controlled by them. They would use the technology if it was exclusive to them.
Me too, I was scared to comment but...I can write lyrics for days, but I do not have the "look" that record labels want, not only that I'm goofy in front of a live camera and I will stutter in front of a crowd and fall all over my lyrics, Suno gave me the voice I needed. I did ghostwriting back in the late 90s, I will never do that again, it's a very sad place, I am happier now barely making a penny off my stuff than I was when people were paying for my lyrics back then
Shame on both of you. Just get a new different dream. Your hack to this dream is cheating and gross.
@YTTraveler777 You are not an artist. You couldn't think of something more creative to say? AI is a tool.
Do you play all the instruments? No!
You can write your lyrics. You don't have to pay a band to record your songs. You can generate the instrumentals and vocals. Your lyrics are not just sitting on paper. It gives bands, vocalists a chance to hear more songs quickly to find the ones they want to perform.
I love this neutral analysis of everything. Thank you!
Thank you for the support. Please join us for our UA-cam Livestream tomorrow.. We host every Wednesday at 5 pm PST. We would love for you to engage there.
What's funny is that some AI "artist" are acting as if they are gonna become super stars, famous or take control of the music industry if the record labels lose the case. Nothing will change.
@@RobHawksbyll The labels will use their own AI to train it on the raw stem/DAW data and hence will create much better music than the current AI can.
"AI artist" now that's an oxymoron.
@@DBecks09 Labels have already started training AI models, and putting provisions for it in artists' contracts
@@troubadour1029”Talentless Wannabes” - TW artists.
Lmao I wanna see these AI "artists" perform a live show. See how much of a super star they are then
I felt the title was a bit clickbaity. So I had ChatGPT pull this episode apart for you all
The title "Major Music AI Lawsuit Update: Suno's Shocking Admission" appears to be somewhat clickbait, as it implies that Suno made a dramatic or unexpected confession. However, the content of the transcription reveals that Suno's so-called "shocking admission" isn't as surprising as the title suggests. The main point is that Suno acknowledged that their AI models trained on data from the open internet, which may include copyrighted material owned by major record labels.
Suno's defense, however, revolves around the argument that training AI on publicly available data (including copyrighted material) is akin to a human learning from listening to music, and they maintain that this doesn't constitute copyright infringement. They also argue that the lawsuit is more about the major record labels trying to stifle competition rather than genuine concerns over copyright infringement.
So, while Suno does admit to using data from the open internet, the admission isn't necessarily "shocking" but rather a predictable part of their legal defense strategy in the ongoing copyright battle. The title of the video may be intended to draw in viewers by suggesting more dramatic content than is actually present.
By the way, I just realized that this false phrase "AI learns just like a child" could be used to activate people's parental instincts, so they will be "oh, then we need to protect our child at all costs" subconsciously. Isn't this a classic lawyer's psychological tool
If music passes away it has to pass away for everyone. I cant be told music is done now, AI took over and then I open instagram and see some new artist on stage performing in front of thousands.
One can like live music and AI music and for different reasons. One can go to concerts for a shared human experience and make personalized music with AI to have songs that speak your words and feelings directly.
How does an AI being trained on existing copyrighted music differ from an artist being "inspired by" those same tracks and artists?
Because...
I don't want to get sued by Nirvana. That's why!
Because AI is just a machine, it does not learn, it does not get inspired , all it can do is copy
@@marksmusicCC Well I've used Suno to create a couple of tracks and they definitely are not copies of any existing work - they are however, obviously inspired by existing works. I don't think "copy" is a valid claim and that's what's at the heart of this lawsuit. It's also worth noting that even when you use Spotify you are copying the track to your local device so copyright owners expect those tracks to be copied as part of a process -- as opposed to copied in lieu of purchase -- hence Suno's defense claim does not weaken their position IMHO.
It is same. The different is, how the AI company get the data (music data or file). They buy it like other people do when they used streaming platform? Or they only use free material from youtube? Or search from torrent site? If they do not use pirate material, it is OK. It is from learning material. The output also subject of the law. Is the output have similarity that can categorize as plagiarism?
@@xjet Suno created the tracks, you didn't. The price of admission to be an artist of any sort is to develop the skills to create it. Want to create music? Do what actual musicians that AI steals from do: LEARN MUSIC THEORY, MUSIC PRODUCTION, AND MUSIC PERFORMANCE. That is "democratic" as it gets. AI is not "inspired" by anything but greed.
Thank you
AI training is not imitating it is the study of the factual aesthetics that create the basis of music. You cannot copyright a fact. If it is a fact that X number of songs use such and such 3 chord progression and guitars have a certain tone when those progressions are played that is a fact. On the level of AI it is doing what the human brain does to learn music, but in an algorithmic manner as opposed to an organic one. The analysis that AI does is akin to the phone book cases and the Google suits regarding links to existing sites which took snippets, which fell under fair use. They are merely being trained via the accumulation of facts and an analysis of facts. Because facts cannot be copyrighted, the RIAA has not one leg to stand on. Just my opinion as a guy with an LLM in Entertainment and Media law.
The AI is not analyzing chord progressions and guitar tones and synthesizing it's own music from scratch using that information, it's using copyrighted audio in it's entirety and categorizing it based on readily available text like tags/genre/etc. If you ask it for a certain genre, it will spit out an amalgam of audio it associates with that genre. Before they tightened the prompts, you could put in niche genres with less training data and have it essentially spit out copyrighted material.
Fundamentally disagree. Comparing the process of AI music "generation" to human inspiration is insulting and completely wrong.
AI replicates music in a way that is virtually indistinguishable from the style and esthetic of the source material. This is because the source material is literally in the models of AI, and is therefore necessarily a constituent part of the final product.
The analogy would be a human being downloading an existing recording whole cloth from spotify, using a DAW to change enough notes to try and avoid a lawsuit, changing the words, but maintaining the style and core components of the original recording - sometimes with even the same exact voice- and then try to claim it's a new piece of music.
That's not musical inspiration or musical creation, it's theft. It's cheap and disgusting.
And to empower some loser this capability to steal music in this way, at the click of a button, is unconscionable
Think about it. Suddenly, regular people with zero musical talent are able to "create" professional sounding music, that sounds just like other artists, and there's no theft behind this?
Very interesting. Please continue the topic and go deeper into this case. What would you do as a Suno lawyer to win this case? Perhaps your ideas will be useful to them. I am on the side of innovation, Suno is a miracle! It understands you like a telepath. At the same time, I hope that there will be reasonable rules for using AI in music. Perhaps, users will be able to create songs with voices and in the style of specific artists, and artists would receive money for it. There should be no choice between live music and AI, they can perfectly coexist.
Thank you for the thoughtful comments. Please join us on our UA-cam Livestreams. We host every Wednesday at 5 pm PST. We would love for you to engage there.
So when do we get the Verdict?????? I have produced over 170 songs on my channel and I need to know ... Do I keep going or throw it all in the trash.
Did you mean AI produced them?
@@natura808no, I believe he means that he collaborated with Ai to co-produce the music.
@@Kaijufruit You mean he typed a prompt
@@natura808 I wrote all the songs. And have to refresh and add prompts to change the mood in certain sections. But yes ultimately there is no doubt that Ai. produced the music. No more than Ai has helped with calculations in the space industry, in automation, for traffic improvement and crime. So yes Ai is definitely creating an assistance of speed in getting to end results. I don't think that Ai music will ever take over specialized artist like Bruno Mars, or Post Malone. Eminem . These artist will always and forever destroy Aii machines with their personalities and authenticity of their music.
Cool. Im gonna become author of 100 books by just writing one line prompt to chatgpt.
Does cover songs, derivative, transformative, etc works get copyright protection? If not, then can AI models use those works to train their models?
AI does not "democratize the arts." If you want to be a musician, or visual artist, or writer, or whatever, there is nothing stopping you but the cost of the equipment and your time. The arts are already "democratic." If you are using AI to "compose music," the AI is creating the music, NOT YOU. You have simply hired technology the same way you would hire a composer from an agency or online labor pool. When you hire a painter to paint your home, you are not a "painter," even if you choose the colors and the pattern. They are doing the work because you lack the skill. If you lack the talent, then you lack the price of admission. People in the arts spend their lives developing their skills. These data douche bags are simply stealing from people in the arts the same way that people who break into museums or record stores or who film concerts for resale might. The courts need to end this madness and admit that it is ALL theft. If you want to create art, learn the skills necessary to create it and stop pretending to be an artist because you can type a few words into a box.
Had the artist never seen another painting? Had the writer never read a book? Had the musician never heard music? I agree, sure, learn the skills and create it yourself. Just where will you learn it and how? From those that came before you. By your logic that makes all creatives thieves.
@@darrenhill123stop spreading misinformation. Human learning and Neural network database training are two completely different things. Don't comment on things you don't understand, it looks ridiculous
Thanks😍, I am also creating test songs, I am also not sure if AI generated songs are disputed or not?
The grandson grandfather thing gives away that they are vultures themselves
And trying to influence the court of public opinion with an emotional play.
its exactly the problem, Major Record Labels ARE making their OWN versions and they want all the power over this tech, its simple AI will take over and is taking over the Music Industry, live with it, use it and enjoy it :) and its simple if they really want to know how and what they need the trained models :) the AI trained models that the AI uses contains all the data, its the same like with GFX the models that AI graphics generators use, they are 5 10 22 gig in size and contain thousands of parameters from images which the AI generator uses to create new images, yes even copyrighted ones from the internet :) thats how they train the model, a new one FLUX uses 12 billion parameters to make new high quality images. With that, think about it, UDIO and SUNO are not crazy, they know how and what when they started this all and for sure knew that the Big Sharks would come after them, do you really think they where and are not prepared to counter them? Perhaps these Record Labels should ask some Expert AI programmers and or Companies that make these things and ask them how things work with AI generators :) this way they perhaps will not look like idiots in the end lmao
I still don't like the fact that now it's like impossible for a new artist to be confirmed legit
The music industry will get nothing for this. What CAN happen is a liscense (required) to access gwenres of music, if such a thing can be agreed upon. What SHOULD happen is a musicians organization forms that creates generic music from all gendres. They can even sing gibberish, and submit to the paralell organization setup to liscense "gig work" to AI platforms. Musch like The Play and App Store work for Apple and Google. Musicians, bands can then write and create generic music for the specific purpose for having it used to train AI. People can money end to end, it's fundable, taxable and a legal avenue to get out of this mess a lot of industries are in. This can ALSO be done with paint/drawing artists.
It's undeniable that the peak of human-composed music was probably reached in the late '80s and early '90s. Since then, it has been on a downward slope. Just listen to the garbage produced in the last decade-it's all driven by marketing hype designed to create global stars out of people like the Kardashians. Give me a break.
The future of music, however, is exciting. Soon, all this commercialized music could be contained in just a few terabytes of self-contained music generation models that can reside on a local PC. People will be able to generate their own music exactly the way they like it. I can tell you that most of the music I listen to now is self-created, and I can't get enough of it. It may sound silly, but it feels like a breath of fresh air-finally free from the garbage that the music industry has been shoving down our throats.
We're no longer subjected to mediocre, talentless so-called artists, made of plastic, who don’t know how to play an instrument, can’t sing, and frankly, can’t even put two artistic sentences together. The joke is on the music industry. And its accolites
I am confused. I personally think the amount of time music can be copywrite protected and owned my major labels is 30 years too long. However it is messed up and arrogant a new music creation platform can dismiss the rules. Everyone else from musicians to labels have to navigate the same playing field.
The Labels may or may not have some kind of case, but it aint copyright infringement, and people who think they do, don't know what copyright infringement is, and what it purpose is.
it is all new territory and new ideas and laws will be designed.
Arguably the training is copyright infringement. Copying includes making a temporary file to train the ai with
One of the issues musicians and copyright will have to deal with is the human ear only hears so much and there are only so many sounds that we find pleasant and as a result we have already past the point of original works.
Nothing is original anymore. It's all copyright or derivative or similar or sampled or a direct note for note repeating of a few to many bars of notes because it's literally all been done already.
If you want "new" music then you have to relax the copyright for samples and similarities.
Is there any chance to argument like this: the constitution says "to promote the progress...", AI is the progress, labels are against the progress, thus copyright can't protect them?
Not at this level. This kind of argument would make sense in a proposal to alter the law itself, as an argument the law isn't fit for its purpose. The US government system has an extreme tendency to sneak in unrelated purposes in the same acts, or counterproductive clauses. E.g. New York Assembly Bill A8955 has an attempt to correct such tampering. In jury trials, there exists a possibility for the jury to override a decision from finding the law unjust, but in another layer of corruption juries that understand this are weeded out.
I have a quick question then dropping in the accusations of monopoly and disproportionate contract terms this would mean they can use discovery process to get examples of this conduct and once it's in the case it's in the public could this be in indirect threat of we know what's going on behind those closed doors and you may take us out if you go ahead with this but be careful what else is in pandora box are you sure you want it all out there
You asked what we thought. Basically it’s already a precedent that it is not infringement whatsoever to use any data of any kind to train an Ai model. Full stop. Period. Done. The rest is pontificating and philosophy in the community, which is reasonable but ill informed. The reality is that based on past cases the record labels have no case, and this is essentially intimidation. The discussion here keeps skipping the actual copyright cases which have already set a precedent on this matter. The reason we have any algorithms whatsoever on any app or search engines or spell checks are based on this fact. If you train a spellchecker on published books that is not infringing. Training an ai model “algorithm” is not infringement. That’s why they are able to harvest all of our data to make all these algorithms for the internet to function as we know it. We may not like it, but it is the way it is.
I totally agree.
Here here
Found Suno's CEO
If AI created music cannot be copyrighted (since it is not human generated), it's value will become unstable as anyone can thus copy or make derivatives without fear of copyright infringement cases or royalty payments.
@@kbuzz123 I generate things people use and buy. Also your argument attacks vocal processors and pitch fixers. Spell checkers. Lol youre not going to stop us. What I create is objectively unique. The Spotify algorithims say so. Nobody cares abut record companies. Nobody EVER cared about artists. Now the writers are here.
Can you tell us in a very few words if we create a song in these platforms in the case large music production companies Winn the case will they claim anything from us in the case that we created a song publish it and probably made a profit from it ?
People are complaining that UA-cam is flagging their original works as somebody else's works. With the millions of songs out there, it will be impossible for there to be truly original works and not some derivative work. If one looks at modern pop rap and bro country, they share the same lyrics, the same "story" and the same type of flow and the same type of instrumentation when it comes to their particular genres. Asking an AI program to make a modern style pop rap is going to produce derivative works because that is what modern pop rap is.
Speaking of UA-cam's filter, just how does that work? Is an AI trained on data given by the labels and can that data be used to make an AI music program?
I believe no music made now is truly original and is just derivative works. Pop acts don't even bother trying and just use samples of past works. Sure these programs will be spitting out derivative works, but that's just because humans aren't as creative as they think they are.
Great comment. This is far from the birth of formulaic music. The industry knows what sells, and have been engineering it for a long time.
At the end of the day, the human interface & creative choices humans make will determine IF a song is worthy of ears hearing it. Popularity is simply a game of relatability. How well you or your collaborator position/collage/ engineer the composition, will determine the “originality” of the artist. it’s when people think it’s ok to say they did it without the collaborative aspect of Ai, even though it’s a “tool”, it also is literally being modeled after human bi-products so naturally there is some bit of humanity in the machine. It’s like we have to tickle it in order for it to loosen up & be playful. Right now it’s in its infant stage & it doesn’t know how to “play”, just how to observe & put blocks together.
Music record labels have always been the Mafia of the industry since the beginning of music, from strong arming artists, under paying them , signing them to such extreme contracts that is practically has caused artist suicides and baring artists for years to even sing or make music.. a violation to me of human rights to perform your god given talents.. now that Ai is out here helping artists and producers get instant reference ideas and so on. as well as allowing your normal every day person to have his or her own songs. Records labels are shit scared of becoming redundant. just like a DVD store.. the world is a new place. and we dont need no Mafia bullshit continuing to ruin the freedom of music..and there are on 8 notes in music people of course there's going to be something sounding like something ..
Training an AI on every song and spitting out variations is NOT the same as someone hearing a metallica song and getting inspired to write a song
Ever heard of a “demo tape”?! they used to make them in the early days of tape recorded music, it’s when an artist or group would play the same song a few different ways to play for others to let them know which version that enjoyed the best. So yeahhh
This also translated into “live recordings” because artists simply got bored with playing the same songs exactly the same way, also connected to the “cover song” & the “remix” but I might be reaching…… for the star dust that we’re apparently all made of….
TBH it pretty much is.
My sister is a professional musician, teaching classical in a variety of stringed instruments, and piano, also at a university co-ordinating a orchestra and vocal students. I myself learned Drums, and Piano when I was young, and in those there is a lot of musical 'Theory' which one could claim is the grammar, however you can not give any computer program the books on Musical Theory, and have them replicate the interpretation and understanding of what is in them. The computer program could only utilize the writ strict 'rules' which always have adaptations, exceptions, interpretations and derivatives to them. (I understand computer programming) I will say that they sent to the 'open internet' these programs to do full data analysis to find those 'personal' interpretations and derivatives and then integrate into the strict limited rules of 'theory'.
not true about "only utilize strict rules", a model can have a 'novelty' parameter which at 0 is strict and at 100(%) fully the oposite