Are AI-Generated Songs Ethical? Let's Talk About It.

[bermudagold]

A lot of music is much less art than product. I think that we use the word artist in something of the same way that Subway uses it. It's not unlike referencing the garbage collector as a "sanitation engineer", part joke, part a genuine attempt to recover some dignity.

for discussion two examples from current pop landscape
Dua Lipa...she was introduced to a Warner Music A&R by a trusted "tastemaker" in his personal circle..."she had nothing" yet in endeavors of music...he said he was willing to sign her before hearing any music at all because of her unique wardrobe, look, charisma, and "energy" when she first walked into the room he was in and commanded everyone's attention....so she debuts with ambassador deals signed for versace, yves st laurent, mulger, and puma

sabrina carpenter was another graduate from the disney kids' tv school, mentored by miley cyrus...her current hit "espresso" is apparently stitched from 4 spice loops, something AI could easily do now...it has deals with Alfred coffee, menotti's coffee, and blank street coffee supposedly before song was released...each deal is exclusive for a geographic region...sabrina posted on all socials for fans to post pics of themselves enjoying her signature coffees for all brands...same with her signature coffee ice cream flavor from Van Leeuwen ice cream

is this music as an artistic endeavor,...or an industrial factory designed for commodity marketing of consumer goods?

[cryophonik]

"Good artists borrow, great artists steal." So, where does that leave AI?

[vurt]

"Good artists borrow, great artists steal." So, where does that leave AI?

isnt theft unethical? so great songs are unethical?
some should be criminal i know that! im looking at you elton

[ROTMetro]

https://deepmind.google/discover/blog/a ... lgorithms/

Even in the earlier days of AI with Adversarial Neural Networks, the process of 'training' is effectively the AI trying things and choosing the parameters that 'work best' over time.

These days, it's far more advanced.

I understand that AI does not switch itself on and think "today I am going to write music", at least not where we have it at the moment. However, dismissing that it cannot learn, or 'experience' things is trying to look at it from a human perception when we don't really understand how 'we' work at those levels.

You don't have to believe me or anything I say. Just listen to those who do know, and follow the direction they are showing as to how AI in research is not the simple 'regurgitate thing from training data' that many seem to think it is:

(also )

AI is 'working out' many things that humans cannot easily, and the way it does this is to 'try' and 'fail a lot' through direct 'experience'.

Modern robotics in industry is often build on this. Digital 'twinning' is a huge industrial research basis that NVidia and others make millions off because it allows robots to 'learn' from doing in a virtual environment. They aren't fed the solution in training data - they iterate and work it out through 'doing'.

I'll not say any more, as clearly my perspective on AI doesn't meld nicely with the way many see the world, but there is far more to what is happening than just 'copying' what it has seen before.

The music models we have access to do not 'learn' anything post training. It is literally fed and re-enforced on 'the solution' during training. You might argue that those doing re-enforcing are making it different from the average as they nudge it's inherent 'averaging all things' functionality, but it's just static averages pushed slightly via training to whatever Suno thinks their customers prefer. Still just an average though, but now an average of what Suno thinks a wanna be AI musician wants, instead of an average of just the music stolen.

[ROTMetro]

"Good artists borrow, great artists steal." So, where does that leave AI?

A weak cliche. But it wasn't intended to justify wholesale destruction of entire industries via stealing others works and generating slop. It was intended as a poetic way to acknowledge that creativity builds on the past, that all of human achievement is build on top of past achievement.

AI is the opposite, it doesn't build anything, what it does do is eat all the future seed corn before it can grow (new artists, authors, entry level workers, everyone just coming up that needs time and space/an income to grow) and stagnates humanity to where we are. AI strip mines other people's works and all future works so that corpos can profit off regurgitation. It's not progress, it's not great art, it's extraction. Extraction by corporations from other peoples works. It's stealing and it's destructive to society long term.

[koalaboy]

Well, buckle up: https://elevenlabs.io/blog/eleven-music-is-here

[ROTMetro]

Yay, the people that enabled the voice cloning boom for phone scammers bring us deeper into their dystopian vision. It's always the worst people pushing this stuff.

The reverb is still crap and a huge AI tell.

[ghettosynth]

"Good artists borrow, great artists steal." So, where does that leave AI?

A weak cliche. But it wasn't intended to justify wholesale destruction of entire industries via stealing others works and generating slop. It was intended as a poetic way to acknowledge that creativity builds on the past, that all of human achievement is build on top of past achievement.

AI is the opposite, it doesn't build anything, what it does do is eat all the future seed corn before it can grow (new artists, authors, entry level workers, everyone just coming up that needs time and space/an income to grow) and stagnates humanity to where we are. AI strip mines other people's works and all future works so that corpos can profit off regurgitation. It's not progress, it's not great art, it's extraction. Extraction by corporations from other peoples works. It's stealing and it's destructive to society long term.

The idea that these music models are just generating “averages” misses what’s actually happening. They don’t blend everything into mush. What they’re doing is sampling from conditional probabilities learned during training. That means they’re choosing what’s likely to come next in a given context, not averaging all possible options. The use of the word average here is a bit of a lay oversimplification. The randomness at each step isn’t noise, it’s added on purpose. That’s where the variation comes from, and it’s what gives the model any chance of producing something that feels new. It’s not unlike how people use creative constraints or randomness to spark ideas. Think of Brian Eno’s Oblique Strategies. You throw something unexpected into the process to break the loop and push things in a direction you wouldn’t have reached by intention alone. A lot of pop songwriting already works this way. You sit with a loop, try ideas until something sticks, toss what doesn’t work, and keep what does. It’s guided trial and error. Once you frame it that way, the gap between human and machine isn’t as wide as people want to believe that it is.

[Michael L]

AI songs are digital and cold, human songs are analog and warm.

[ROTMetro]

The idea that these music models are just generating “averages” misses what’s actually happening.

"it's an engine to give the most 'likely' next step (totally not the average next step, just the most likely one based on weighted averages from the stolen training data) at each point by using weighted...measurements (not averages)"

A human mind with agency is nothing like a machine iterating on weighted past behaviors. You have nothing transformative. You have machine processing of huge amounts of data. If it was transformative you would get novelty even from sparse input. If I train on 1 song, I will not get transformative novelty. It’s operating within the latent space of what it’s already seen. There's no spark. No surprise. No intention. What can AI do with one song? Not much. Because there's no 'there' there.

[ghettosynth]

The idea that these music models are just generating “averages” misses what’s actually happening.

"it's an engine to give the most 'likely' next step (totally not the average next step, just the most likely one based on weighted averages from the stolen training data) at each point by using weighted...measurements (not averages)"

Not completely, "a likely" next step. If it were always the most likely they would create deterministic outputs relative to input prompts. That's the function of temperature and friends in terms of guiding the process.

The use of "stolen" here is loaded. Trained from is not the same as sampled and the courts are still working on how this fits into the world.

A human mind with agency is nothing like a machine iterating on weighted past behaviors.

No, a human with agency is very much like a machine iterating on weighted past behaviors. Just because the training and inference steps are more tightly integrated and the leaps from one idea to the next function in a different domain doesn't remove the similarity.

You have nothing transformative.

Hard disagree. Whether you like the content or not "I glued my balls to my butthole again" is transformative. If not written by AI it would be subject to copyright protection. If you mean a deeper meaning of transformative than the copyright office, well, I don't think that Achy Breaky Heart meets that standard when compared to Tulsa Time, as just one example.

You have machine processing of huge amounts of data. If it was transformative you would get novelty even from sparse input.

First, this is your cherry picked definition of transformative, and second, it's naive to think that any human capable of producing a transformative music work, by almost any definition of transformative, is functioning on just a single input of any particular song.

[eassae]

https://www.instagram.com/p/DND7QoOMwuh/

[ROTMetro]

Way too many quotes I don't have the brain energy to parse it all up pretty..

Human's, on their own, spontaneously came up with music, all across the entire planet. People can vary easily transform a single song without other reference. Spend time with any small child. And there's zero call to call me naive, make your point not personal attacks.

If I sell a product, and that products is only useful because it had access to music, I have stolen from that music. Period. If the AI tool would be less useful as a product without taking from other's work, then the AI tool is....taking from other's works. That fact that they trained on those works shows they are necessary for the product, so the product is derivative. If it wasn't core to their product they wouldn't have trained on copyrighted music.

Not completely, "a likely" next step. If it were always the most likely they would create deterministic outputs relative to input prompts. That's the function of temperature and friends in terms of guiding the process.

The use of "stolen" here is loaded. Trained from is not the same as sampled and the courts are still working on how this fits into the world.

A human mind with agency is nothing like a machine iterating on weighted past behaviors.

No, a human with agency is very much like a machine iterating on weighted past behaviors. Just because the training and inference steps are more tightly integrated and the leaps from one idea to the next function in a different domain doesn't remove the similarity.

You have nothing transformative.

Hard disagree. Whether you like the content or not "I glued my balls to my butthole again" is transformative. If not written by AI it would be subject to copyright protection. If you mean a deeper meaning of transformative than the copyright office, well, I don't think that Achy Breaky Heart meets that standard when compared to Tulsa Time, as just one example.

You have machine processing of huge amounts of data. If it was transformative you would get novelty even from sparse input.

First, this is your cherry picked definition of transformative, and second, it's naive to think that any human capable of producing a transformative music work, by almost any definition of transformative, is functioning on just a single input of any particular song.

[ghettosynth]

Human's, on their own, spontaneously came up with music, all across the entire planet.

Define spontaneously. Any human interaction with music is informed by past experiences, full stop. We listen and learn intuitively, or formerly through explicit study, all of our lives. At that point that we are "spontaneously" creating music, our output is based upon the sum total of our interaction with the music of others at whatever level we are able to project our internal music onto whatever sound making device that we choose to use.

No human accomplishes much of anything that is not build on the shoulders of others.

People can vary easily transform a single song without other reference. Spend time with any small child. And there's zero call to call me naive, make your point not personal attacks.

A small child that can understand language enough to understand the task has already trained their inference engine significantly from the works of others. Whatever translation that the child performs, from incompetent to gifted, is a function of the knowledge gained from being exposed to the world. You are trying to erase this as if it doesn't matter. A newborn will just look at you while they fill their diaper.

Your idea was either disingenuous or naive. I should have to make the point that humans learn from other humans and their environment, that is obvious, and, as you assert, should be clear if you've ever spent any time around small children. So you are either articulating a naive idea, or a disingenuous one. Naive here is used in the clinical sense about your statement. The use of words that are incorrectly perceived as offensive does not constitute an ad-hominem attack. The perception that they do is, wait for it, naive.

If I sell a product, and that products is only useful because it had access to music, I have stolen from that music. Period.

Again, your attempt to recast training from as sampling, or worse, theft, is loaded because you think that it helps to make your point. Every time that you listen to someone else's work and then subconsciously create similar patterns in your own work, by your definition, you are stealing. To use your own example, the child that transforms Mary Had A Little Lamb into Mary Had A Little Scam is stealing from the original. The transformation demands that the original is copied, precisely because, the granularity of pattern decomposition available to the child is quite limited.

So, in that sense, humans, in your language, "steal" more than AI does. I know that bugs you, but failing to accept it when you yourself constructed the frame of analysis is disingenuous.

If the AI tool would be less useful as a product without taking from other's work, then the AI tool is....taking from other's works.

Again, it's not taking unless you are also taking. This should be obvious to you. We even have words like genre and style to capture in broad stokes how humans copy from other humans. In fact, this often serves to gatekeep work that is too transformative. "That's not techno, that's something else, get out of our clique with your non-techno nonsense." We expect our music to sound like other music, just not exactly, that's how we build patterns of what we like, and, wait for it, those patterns guide us when we are creating "new" work.

That fact that they trained on those works shows they are necessary for the product, so the product is derivative. If it wasn't core to their product they wouldn't have trained on copyrighted music.

Yes, the training is necessary. However, what seems to bug you is that it's not in your purview to decided whether that's copyright infringement or not. It's not a copy, it's a listen. Just as humans listen and update their own inference engine.

The only reason that the training works, however, is because humans copy other humans leading to patterns that can be statistically evaluated. If every work was truly unique there would be no patterns from which to derive statistical relationships between language and music tokens.

Progress always induces anger, especially among the gatekeepers. The gatekeepers will lose here. The models will not be discarded, that would simply give the AI advantage to foreign nations.

[ROTMetro]

Bro stop. So you think that what, aliens gave us the first music? And we just copied from that forever? No, music was invented all over the world, spontaneously, by human beings. Because we are amazing creative beings, because we can enjoy things. Because we like to share experiences. Because human's feel love and heartbreak. Because humans live, and die. Because of sunny summer days, and traumatic loss and grief. Because of tons of distinctly human reasons. Lot's of music is building, but lots is also spontaneous creation and a need to share the deeply personal. AI CAN'T do either bit. A trained AI doesn't learn after the fact, it's model doesn't grow, it doesn't build. It spits out regurgitated solutions it got form other's work. It's a static for profit product. That is why new models come out, because they are the 'grown' dataset. AI doesn't build. And it doesn't spontaneously create.

I like this quote today from the current HN discussion on GPT5:

"Human intelligence is markedly different from LLMs: it requires far fewer examples to train on, and generalizes way better. Whereas LLMs tend to regurgitate solutions to solved problems, where the solutions tend to be well-published in training data."

But also unless you are a billionaire or a tech bro, humans aren't products. AI is a product. What we are talking about is a commercial product, scaled well past one persons' output, designed to be sold to replace entire industries of workers. A commercial product is held to a different standard than a single human, especially when each human takes decades or personal work to grow, while an AI can be massively scaled, and only exists to be a sold for profit commercial product. The two are not the same, and copyright standards and understanding should not be the same for the two.

If your commercial product is using others' works, you are taking from them. It doesn't matter you have an intermediate 'trained model' between the other persons work and your final commercial product. Your product is derived EXPLICITLY from the work of others (copyright covers 'derivative' works. The entire AI product is a derivative work, it would exist without the original source data). If you are deriving a product to sell from other people's work, if you're product wouldn't exist without it actually ingesting and processing someone else's, you are taking from others. AI is a product, not a person. And the part that makes the product useful is the dataset derived from...ingesting other peoples stolen work.

I think AI is cool. I think it could be powerful. But we need to be honest that it is a product EXPLICITLY derived from the work of others. It's not a human being, we don't get to say 'because humans can learn from stuff, making a product that is dependant on and derived from the work of others is ok'.

You are an AI zealot and I'm just a music loving software developer with an expensive synth hobby, there's no point for further talk between us as we are just talking past each other now. Instead of a true social analysis on this that a real society would have, this will play out in court and politics, and the big money tech bro's will buy their way to success, and the laws and civil standards developed over centuries to empower people sharing knowledge (copyright law) will be labeled as 'gatekeeping' because it's inconvenient to a commercial product's profitability. People who think theft is wrong will be called 'angry, irrational', because stopping that theft and making the tech bro's pay what they owe under the law would impact a commercial products (not a person, but a commercial product intended to make BILLIONs from it's training set) profitability. People that think 'maybe we should have societal impact' level discussions on something that has such far societal impacts will be called 'gatekeeper, reactionary, angry'. F' off bro. Again you negatively and unfairly label. You are malicious in your replies.

And we will all have smaller, greyer lives because of it.

This is the second time you have reduced to bs name calling (gatekeeping, angry, blah f'n blah) so there's no need for us to respond to each other and you're blocked. Neither of us owes the other any of our attention nor thought simply because we posted on KVR, and I'm done giving you any of mine since you've chosen multiple times to name call.