Generative AI puts interests of Google vs the interests of YouTube

[WillyW]

Summary

 A Yutube user created a generative AI song with sources from Drake and The Weekend. There was some drama with the song being taken down across various different sites including Apple Music and Spotify, but there was further drama with Youtube. The issue at hand is that a DMCA takedown notice was only applicable to the song because of a specific sample in the song. If the user removed that sample it would pit Youtube's interest in direct contrast to Google's generative AI businesses. This will become complicated as the label that represents Drake and The Weekend, UMG has publicly stated that they do not view generative AI as fair use, but Google themselves have said in the past that generative AI products are fair use.

 

Quotes

Quote

The first legal problem with using AI to make a song with vocals that sound like they’re from Drake is that the final product isn’t a copy of anything. Copyright law is very much based on the idea of making copies — a sample is a copy, as is an interpolation of a melody. Music copyright in particular has been getting aggressively expansive in the streaming age, but it’s still all based on copies of actual songs. Fake Drake isn’t a copy of any song in the Drake catalog, so there’s just no dead-ahead copyright claim to make. There’s no copy.

Quote

The problem is that Google, Microsoft, StabilityAI, and every other AI company are all claiming that those training copies are fair use — and by “fair” they do not mean “fair as determined by an argument in an internet comments section,” but “fair” as in “fair as determined by a court on a case-by-case application of 17 United States Code §107 which lays out a four-factor test for fair use that is as contentious and unpredictable as anything in American political life.”

Quote

What do you do?

• If Google agrees with Universal that AI-generated music is an impermissible derivative work based on the unauthorized copying of training data, and that YouTube should pull down songs that labels flag for sounding like their artists, it undercuts its own fair use argument for Bard and every other generative AI product it makes — it undercuts the future of the company itself.
• If Google disagrees with Universal and says AI-generated music should stay up because merely training an AI with existing works is fair use, it protects its own AI efforts and the future of the company, but probably triggers a bunch of future lawsuits from Universal and potentially other labels, and certainly risks losing access to Universal’s music on YouTube, which puts YouTube at risk.

 

My thoughts

In the past record labels have been quite litigious and given that the generative AI business is also predicted to be quite profitable for the tech giants, at a minimum there will be some upcoming changes in Copyright law. In the past lobbyist groups for major rights holders groups like record labels have successfully obtained provisions in legislation and trade agreements such as the TPP. It really is anyone's guess who will have the bigger say, but personally any time the RIAA runs into problems I am generally happy given the underhanded tactics that they have tried to protect their profits whilst not increasing the profit from artists. The Recording industry is quite an exploitative industry if you are not a big name.

 

Sources

https://www.theverge.com/2023/4/19/23689879/ai-drake-song-google-youtube-fair-use

[johnno23]

its always about money......anyone remember when Apple released Garage band ?

Some labels got real hot under the collar due to samples being used at complete loggerheads with many artists at the time that thought kids in their bedrooms sampling and creating something new was awesome..........thats life sadly.

[Brooksie359]

Honestly I feel like so long as people aren't making money off of these things it should be fine. The problem becomes when people take someone's voice through AI to create music and then try and profit off of it. At that point you are basically stealing someone's likeness through their voice without permission which tbh is just a scummy thing to do. Now if it's just a fan made music that isn't for profit I don't see it as much of an issue. I mean I have seen a bunch of stupid AI covers of songs that are very entertaining and it would be a shame if they were made illegal even if they weren't monetized. 

[johnno23]

Agreed but companies so often bend rules to suit their own bottom line....how often has an actors voice been sampled for a dialogue in a movie but the actor has passed away? 

Did the movie company pay the relatives or simply decide they owned the rights as the dialogue was sample from something they said elsewhere for them ??

Either way AI is gonna put millions of people out of work. 

Who needs to pay for an expensive fashion model the photograper the studio etc etc when a fashion magazine can simply give the AI photos of the garmens drop in the prompts for how the model should look the type of pose etc and minutes later done and dusted.

Advertising will sieze upon it as who needs an actor or actress to promote a toothpaste anymore when AI will do it in moments? Celebrities will benefit as they are often used for the recognition factor but all the other simple no name actors etc are so easily replaced.

As for Music it is simply hopeless as in the style of is contestable and when people can simply talk and have AI transform their voice to a melody that is in tune and on beat there is no limiting factor.

We are living in one of those moments in time where it is both amazing and scary at the same time.

[Brooksie359]

9 minutes ago, johnno23 said:

Agreed but companies so often bend rules to suit their own bottom line....how often has an actors voice been sampled for a dialogue in a movie but the actor has passed away? 

Did the movie company pay the relatives or simply decide they owned the rights as the dialogue was sample from something they said elsewhere for them ??

Either way AI is gonna put millions of people out of work. 

Who needs to pay for an expensive fashion model the photograper the studio etc etc when a fashion magazine can simply give the AI photos of the garmens drop in the prompts for how the model should look the type of pose etc and minutes later done and dusted.

Advertising will sieze upon it as who needs an actor or actress to promote a toothpaste anymore when AI will do it in moments? Celebrities will benefit as they are often used for the recognition factor but all the other simple no name actors etc are so easily replaced.

As for Music it is simply hopeless as in the style of is contestable and when people can simply talk and have AI transform their voice to a melody that is in tune and on beat there is no limiting factor.

We are living in one of those moments in time where it is both amazing and scary at the same time.

That being said music even when on beat isn't always good. There will still be good songs and bad songs. I doubt we will see something where everyone is now good at making songs. If that was the case then there would be a ton of amazing EDM songs which don't require vocals but even with EDM there are definitely people who are objectively amazing at making EDM songs that can't be replicated at this point. 

[Quackers101]

4 hours ago, Brooksie359 said:

That being said music even when on beat isn't always good. There will still be good songs and bad songs. I doubt we will see something where everyone is now good at making songs. If that was the case then there would be a ton of amazing EDM songs which don't require vocals but even with EDM there are definitely people who are objectively amazing at making EDM songs that can't be replicated at this point. 

it will just be more drowned in content, like the new releases on the steam page. So projects with more work into them, will drown more and benefit less, unless given some sort of push by the system or is from well known names/brands or had a good PR push.

 

For new or smaller people this can suck, but if one are able to make the tools help you and cost less in the losses of visibility.

 

To a future where generative AI has to be "open source" and act like there is no real "ownership" over its content, only what the creator does "privately" or adds to the content, not generated due to the issues of noise and what can be "public assets". I guess.

Edited April 21, 2023 by Quackers101

[Kisai]

6 hours ago, Brooksie359 said:

Honestly I feel like so long as people aren't making money off of these things it should be fine.

Which you can't do because opt-ing out of the ads on youtube isn't possible.

6 hours ago, Brooksie359 said:

The problem becomes when people take someone's voice through AI to create music and then try and profit off of it. At that point you are basically stealing someone's likeness through their voice without permission which tbh is just a scummy thing to do. Now if it's just a fan made music that isn't for profit I don't see it as much of an issue. I mean I have seen a bunch of stupid AI covers of songs that are very entertaining and it would be a shame if they were made illegal even if they weren't monetized. 

 

Here's my overall opinion. 

 

Training the AI is fair use. Creating a commercial work INTENDED to imitate the training material is NOT. 

 

Eg, I could train an AI to sound like Robin Williams (who is DEAD), and another AI to sound like Gary Owens (Who is also DEAD) but I can not take the resulting AI's and call them Robin Williams and Gary Owns sing a duet. It's impossible to get their permissions, it's impossible to get them to create material explicitly to train an AI. So the only option is to lean on the fair use and go "I have this AI trained called FunE-Man that can sound like Robin or Gary, but is not either of them." 

 

Then I could take that AI trained on them, and train a third AI, that only learns from the AI-generated voice. The result is a voice that has not been exposed to the original training voices, and operates under the same clean room logic that allowed COMPAQ to clone the IBM bios, and Accolade and EA to create clone cartridges for the Mega Drive, and Bleem to create the PS1 emulator at Sony's expense.

So, you could in theory clone a singer's voice, but you can not produce a low-effort cover using the original singers voice against a different mix of the song. It would damage the commercial value for the original song. After all, this is exactly why singers re-record their own songs. The music industry sucks and the way a company can own "the recording" without the singer having a say, is the same way a singer's voice can be cloned and continued to be use without their permission. The recording industry has to give up that claim on the recordings if they want to go after the AI voices. They can't have it both ways.

 

[Brooksie359]

3 hours ago, Kisai said:

Which you can't do because opt-ing out of the ads on youtube isn't possible.

 

Here's my overall opinion. 

 

Training the AI is fair use. Creating a commercial work INTENDED to imitate the training material is NOT. 

 

Eg, I could train an AI to sound like Robin Williams (who is DEAD), and another AI to sound like Gary Owens (Who is also DEAD) but I can not take the resulting AI's and call them Robin Williams and Gary Owns sing a duet. It's impossible to get their permissions, it's impossible to get them to create material explicitly to train an AI. So the only option is to lean on the fair use and go "I have this AI trained called FunE-Man that can sound like Robin or Gary, but is not either of them." 

 

Then I could take that AI trained on them, and train a third AI, that only learns from the AI-generated voice. The result is a voice that has not been exposed to the original training voices, and operates under the same clean room logic that allowed COMPAQ to clone the IBM bios, and Accolade and EA to create clone cartridges for the Mega Drive, and Bleem to create the PS1 emulator at Sony's expense.

So, you could in theory clone a singer's voice, but you can not produce a low-effort cover using the original singers voice against a different mix of the song. It would damage the commercial value for the original song. After all, this is exactly why singers re-record their own songs. The music industry sucks and the way a company can own "the recording" without the singer having a say, is the same way a singer's voice can be cloned and continued to be use without their permission. The recording industry has to give up that claim on the recordings if they want to go after the AI voices. They can't have it both ways.

 

Just swear a bunch of times before the song starts. The video will get demonitized based on youtube guidelines. 

[mr moose]

It probably only needs one good case to go through all the courts to deem whether or not the sound of your voice is akin to IP.  It is the thing that's makes a singer worth what they are,  being able to mimic their tonal qualities may well be interpreted by the courts as a sonic trademark infringement.

 

 

[Doobeedoo]

Or Alphabet hmm. But whatever. 

[WolframaticAlpha]

15 hours ago, Kisai said:

Snip 

If Spotify releases an AI that allows people to enter in the prompts and create music, which includes the ability to make music in the style of someone, will it be infringement? Really just asking.

[Kisai]

16 hours ago, Brooksie359 said:

Just swear a bunch of times before the song starts. The video will get demonitized based on youtube guidelines. 

That was already walked back, but that is one way people previously were forcing contentID claims on their videos to be worthless.

 

10 hours ago, mr moose said:

It probably only needs one good case to go through all the courts to deem whether or not the sound of your voice is akin to IP.  It is the thing that's makes a singer worth what they are,  being able to mimic their tonal qualities may well be interpreted by the courts as a sonic trademark infringement.

 

 

 

I'm split on how I want to see this shake out.

 

On one hand, people who are no longer alive, you can not get permission from, but the recording industry (also the film industry) has enough raw material to not only train an AI to duplicate 90% of how someone sounds, if they've appeared in one film, or recorded one complete CD's worth of audio (you only need about an hour to get reasonably close, ignore zero-shot TTS stuff, that stuff is all garbage and closer to "AI doing an impression of someone's voice", fairly easy to tell.) Once you have an hours worth, you need to find all the key phrases they are known for, so the AI can learn how to say those things. We as yet have no way to tell AI's how to emote, because that's a limitation of present AI's only being trained on how to "speak" like a narrator, instead of having a conversation. AI's don't scream or teakettle.

 

On the other hand, it feels a lot like grave-robbing to use someone who is dead's voice. As much as I may like an actor or a singer, we tend to look at impressions of said actors and singers as a bit in bad taste (eg Elvis) but a lot of people are paid to do it, and an AI doing this is exactly the same in that regard. Likewise the recording companies continuing to sell records/CD's of dead singers, and even alive ones, against the family's wishes.

 

Like, I know which way the recording and film industry wants this to go, they want to be able to create AI actors and singers and use them forever, just like their recordings of existing people. However voice actors, stage actors, film actors, and singers don't want this, and unions are pretty much all in agreement that that it would be bad for them to be replaced by AI. So I don't see the recording industry fighting that hard here because if they do argue that AI can't be used, then they are arguing that they can't use their own recordings if the person's voice who is on it revokes or refuses to have that recording out there any more.

 

4 hours ago, WolframaticAlpha said:

If Spotify releases an AI that allows people to enter in the prompts and create music, which includes the ability to make music in the style of someone, will it be infringement? Really just asking.

Right now, anything generated by AI is de-facto public domain. So if you said "make a musical composition for an anime opening in the style of BTS" it would likely make something that musically sounds genre similar, but not recognizably as BTS.

 

What the recording companies are afraid of is someone going "I would walk 500 miles", "Hatsune Miku", and getting a cover of the song, that they are not being paid for.

BTW, this miku video has appeared and disappeared at times for exactly the same reason. It disappears because some rights holder claims it, and then it re-appears later after some time. It's 15 years old. The point of the video is not that "it's a cover of a song", but that the Vocaloid "singer" is an instrument.

 

For the recording industry it's only about money. So even someone going "isn't this kinda cool" crosses the line.

 

For artists (be it musical works or visual works) they are very scared that they will be replaced by AI and their name will be used with the AI, thus creating a brand confusion that these "AI" versions of their performances are them, even crappy ones.

 

Here's the nightmare scenario...

 

Let's say I create an AI that can duplicate a singer's voice from just one non-singing use of their voice, eg an interview on one of the late night shows, where there is no backing music or audience noise. I then go through the list of thousands of existing karaoke midi files that were made in in the 90's, and have this "AI" version of them produce a cover of every single song out there, even paying the "fee" to have the cover.

 

Now, you laugh, but this is what people do already. Only they've been doing them as 8-bit covers, and presumably with spotify's song-to-midi tool. I'm not saying that someone isn't capable of putting out one-song-per-day, but it seems that this is something more likely done with said tool than an actual person listening to the original song and deciding what 8-bit sound goes with each beat. One of these channels has 4000 videos, and has an itunes store.

 

Now remember the issue everyone has with youtube flagging covers of songs on videos? When you play legend of the zelda, and some weirdo in europe has a "Cover" of the song that is them just adding nonsense lyrics to it? And that is what you get a claim from?

 

Those 8-bit covers likely are doing the same thing.

 

The nightmare scenario already exists without the AI generating singing itself. It's just going to get whole lot worse, because people will put up these covers and them claim uses of the originals.

 

There is an argument to be made that "tailored for you" is good, but "machine generated spam" is not. Maybe I want to listen to William Shatner rap like Eminem, for funsies. AI can already do that, because Rap tends to align closer with how TTS actually sounds in the first place. But present AI systems were not designed for singing, because no singer in their right mind would ever let you use their voice for an AI. Even Vocaloid's that exist? They used voice actors, not singers. Yet, even though we've had Vocaloid instruments since like 2007 (Vocaloid 2), (Vocaloid 1 is 2003), it didn't take off until Miku, and despite there being 20 years of this stuff out there, there isn't a pile of Miku-covers-everything supplanting Japanese and English music. It's just not the thing people want to listen to.

 

And I feel that if AI singers get over the hump of "good enough for a cover" to where original music starts being made with them, well we fall right back into the argument that AI generated material is public domain.

This is a track someone made, using the Japanese TTS google uses (at the time.) It was then turned into a song. So let's say that Google changes the TTS voice used later and this original voice disappears. It's then no longer possible to create this song, only the person who originally made it and has the recording can. However, remember there are like two parts to copyright. There is the musical work (the piano roll), and the lyrics. The reason anyone can make a cover is because of how the mechanical license is split. So is this track in the public domain? Probably not. The musical composition is original, but it's based off how the google TTS voice sounds. The TTS audio likely does not qualify for copyright protection. So would it be legally permissible to re-produce this track (how it's done is literately in the video), or is there some part that is actually copyrightable?

 

It's things like this that ask the "at what point do we consider something original" question. My opinion is that the complete composition probably does count, but if someone were to make a cover of it, they probably could not argue they own enough of it to claim full ownership of it since it was generated by AI. If the AI was not there, the song would not have been.

 

 

 

[Brooksie359]

10 hours ago, mr moose said:

It probably only needs one good case to go through all the courts to deem whether or not the sound of your voice is akin to IP.  It is the thing that's makes a singer worth what they are,  being able to mimic their tonal qualities may well be interpreted by the courts as a sonic trademark infringement.

 

 

I think this would fall under likness clause. I mean if you look at thr definition of likeness voice is included. Also likeness is something that is protected so I think you could easily get them so long as it's for commercial gain. 

[LAwLz]

22 hours ago, mr moose said:

It probably only needs one good case to go through all the courts to deem whether or not the sound of your voice is akin to IP.  It is the thing that's makes a singer worth what they are,  being able to mimic their tonal qualities may well be interpreted by the courts as a sonic trademark infringement.

The problem with that logic is that the same should apply to singers that sounds alike as well then. If you are capable of owning the sound of your voice, then other people mimicking it should be infringing your IP as well. 

 

Are you able to singing and sounding like a famous artist? Sorry, but if you release a track of you singing they are sue you for copying their IP. 

[mr moose]

2 hours ago, LAwLz said:

The problem with that logic is that the same should apply to singers that sounds alike as well then. If you are capable of owning the sound of your voice, then other people mimicking it should be infringing your IP as well. 

 

Are you able to singing and sounding like a famous artist? Sorry, but if you release a track of you singing they are sue you for copying their IP. 

 

I don't see why it has to,  One is a direct attempt to earn off the back of another artists while someone sounding the same still has to earn their place,  I'm thinking bands like Greta van fleet, that even though they are clearly on the led zepplin train, they still have to work to produce their music, get it out there, be good enough for a contract or to fill venues and repeat the efforts.  Not only that, but they don't shy or pretend they aren't heavily influenced by zepplin.   Training (in contrast) AI does not take anywhere near the physical effort, cost or risk.    Essentially one requires a physical skill/talent, the other just requires you know how to use AI.

 

For those living under a rock, here's what they sound like and tell me they haven't embodied everything that makes zepplin zepplin.

 

 

[wanderingfool2]

It's really going to be a case by case basis really.

 

At it's heart, I think that there isn't a problem using someone's voice as a training set (although they will probably lock things down with contract law or something).

 

If it gets wholeheartedly decided that using an AI for this is against copyright then that would also open the door to lawsuits against impressionists.  I think the key would be stating that it's not really the person though.

 

Then again, there are some people who sound so much like other artists as well anyways.

[HenrySalayne]

50 minutes ago, wanderingfool2 said:

At it's heart, I think that there isn't a problem using someone's voice as a training set (although they will probably lock things down with contract law or something).

Yes, there is.

51 minutes ago, wanderingfool2 said:

If it gets wholeheartedly decided that using an AI for this is against copyright then that would also open the door to lawsuits against impressionists.  I think the key would be stating that it's not really the person though.

Why?

 

[LAwLz]

2 hours ago, mr moose said:

I don't see why it has to,  One is a direct attempt to earn off the back of another artists while someone sounding the same still has to earn their place,  I'm thinking bands like Greta van fleet, that even though they are clearly on the led zepplin train, they still have to work to produce their music, get it out there, be good enough for a contract or to fill venues and repeat the efforts.  Not only that, but they don't shy or pretend they aren't heavily influenced by zepplin.   Training (in contrast) AI does not take anywhere near the physical effort, cost or risk.    Essentially one requires a physical skill/talent, the other just requires you know how to use AI.

 

For those living under a rock, here's what they sound like and tell me they haven't embodied everything that makes zepplin zepplin.

If the sound of someone's voice is a copyrightable concept then it will apply to everything, not just AI-generated sounds.

This whole "it is only copyright protected when someone tries to use AI to generate it, but not copyrightable in other situations" needs to stop. Either something is copyrightable or it isn't. 

 

Are people allowed to copy the sound of someone's voice? Yes or no. The method of copyright does not matter, because it has never mattered for copyright law. It is simply not part of copyright law. Copyright law is about protecting an idea or concept itself. The method of how someone breaks copyright does not affect whether or not the copyrighted thing is copyright-able to begin with.

 

 

In order for people to find the AI "guilty" of copyright infringement, we have to first establish that someone's voice is copyrightable. If someone's voice is copyrightable, then it is protected from all methods of copying, not just AI, which would include human imitations.

[HenrySalayne]

11 minutes ago, LAwLz said:

If the sound of someone's voice is a copyrightable concept then it will apply to everything, not just AI-generated sounds.

This whole "it is only copyright protected when someone tries to use AI to generate it, but not copyrightable in other situations" needs to stop. Either something is copyrightable or it isn't. 

You don't understand the problem. If you use the works of an artist in machine learning, you are incorporating their IP into the output of your generator.

 

If you are training your AI without permission on Kurt Cobain you would infringe copyright. If your AI is only trained on Kurt Cobain soundalikes (with their permission) and you would prompt it to output a song in the style of Kurt Cobain - it would be completely fine.

 

This is nothing new.

29 minutes ago, LAwLz said:

In order for people to find the AI "guilty" of copyright infringement, we have to first establish that someone's voice is copyrightable. If someone's voice is copyrightable, then it is protected from all methods of copying, not just AI, which would include human imitations.

That's not how it works.

 

[Brooksie359]

8 hours ago, LAwLz said:

The problem with that logic is that the same should apply to singers that sounds alike as well then. If you are capable of owning the sound of your voice, then other people mimicking it should be infringing your IP as well. 

 

Are you able to singing and sounding like a famous artist? Sorry, but if you release a track of you singing they are sue you for copying their IP. 

Sounds like is not the same especially if they are an person singing. There are already likeness rights that protect against the use of someone's voice especially if it in conjunction with say their name. I would imagine we could easily modify these rights to fight against AI use of someone's likeness but in a more targeted way like using someone's likeness to train AI would be infringement if it was used for commercial gain. 

[LAwLz]

5 hours ago, HenrySalayne said:

If you use the works of an artist in machine learning, you are incorporating their IP into the output of your generator.

No you're not. Zero data from the original input is included in the output. If you think that it is then I suggest you read some research papers on how these programs work. They don't incorporate data from the inputs in the outputs. 

 

 

5 hours ago, HenrySalayne said:

If you are training your AI without permission on Kurt Cobain you would infringe copyright.

Where is your source for this statement? This has never been proven so I'd appreciate if you stopped saying things are a certain way when you have no idea if they are that way. 

 

 

5 hours ago, HenrySalayne said:

That's not how it works.

Yes, it is. Copyright protect works. If we agree that the sound of someone's voice is copyrightable then the protection of that voice, according to copyright law, is protected. It doesn't matter how someone creates a derivative work since it would be the voice itself that was protected. If the court rules that it's the voice that is protected by copyright, then imitations done by humans will be copyright infringement as well. 

 

What could be argued is that training AI on copyrighted material isn't fair use, but that's a VERY different thing from saying a voice can be copyright protected (which is what mr moose said).

[HenrySalayne]

7 minutes ago, LAwLz said:

No you're not. Zero data from the original input is included in the output. If you think that it is then I suggest you read some research papers on how these programs work. They don't incorporate data from the inputs in the outputs. 

Ah, yes, the "read some research papers" argument...

*cough*

 

You can extract training data from diffusion models. Coincident? No, that's how training works. Just because it's probabilistic doesn't mean you're algorithm creates something "original".

 

13 minutes ago, LAwLz said:

Where is your source for this statement? This has never been proven so I'd appreciate if you stopped saying things are a certain way when you have no idea if they are that way. 

Copyright law?

If you can show that the original works are an integral part of the "generator" (which they are - see example above), it will become really hard to argue against it. And then there are personal rights.

 

28 minutes ago, LAwLz said:

What could be argued is that training AI on copyrighted material isn't fair use, but that's a VERY different thing from saying a voice can be copyright protected (which is what mr moose said).

The voice itself is not copyrighted - but it doesn't need to be.

Recordings of your voice and pictures of yourself are generally personal data and your personal property (EU). You can decide what other people can do with it. There are some exceptions for public figures, but I doubt "using them to train my AI" is in the public interest and therefore part of these exceptions.

That's why a "deepfake" was already outlawed way before deepfakes were a thing. No new laws were necessary.

[Kisai]

3 hours ago, HenrySalayne said:

 

 

Copyright law?

If you can show that the original works are an integral part of the "generator" (which they are - see example above), it will become really hard to argue against it. And then there are personal rights.

 

The voice itself is not copyrighted - but it doesn't need to be.

Recordings of your voice and pictures of yourself are generally personal data and your personal property (EU). You can decide what other people can do with it. There are some exceptions for public figures, but I doubt "using them to train my AI" is in the public interest and therefore part of these exceptions.

That's why a "deepfake" was already outlawed way before deepfakes were a thing. No new laws were necessary.

You are conflating two separate issues

 

If I take a commercial recording, regardless of it's source (eg film, music, podcast, talkshow, etc) and use about 1% of it to train an AI. That is "fair use". The actual training of it can always be justified as fair use. It has zero impact on the commercial work.

 

If I take someones voicemail greeting and zero-shot train that against an AI model I already have, and then go use it to impersonate them in some way for my personal gain. That's not fair use. That voice mail greeting was only intended for people who are phoning that person, using it in any other shape falls under the misuse of private data.

 

Now if I train someone's voice with the goal of commercializing the result. That is where it's unclear. Even if the input side is all public domain (which is what LibriTTS is) there isn't individual permission from the people's voices made those recordings, dead or alive, to create a commercial "voice" from their voice. Their recordings were only ever intended to be used as audiobook narration. Not to become a singer. If my AI takes their voice and is only used to narrate material I already own or create, that has zero impact on the commercial value of those librivox audiobooks they already made. It's just not nice to go "hey yeah, that AI voice I made, it's really (NAME), a living person." It's not them. It's an AI voice, and using their name CAN damage them.

 

Like with the deepfake example. If someone who made an audiobook, is unaware that their voice has been used to train an AI (which is kind of what happened with the TikTok TTS voice), and it turns out some commercial service is getting use out of it. Wouldn't you feel absolutely robbed? Or when kids on Tiktok use the voice to create profanity laden videos, how utterly humiliating when you would never say those things.

 

Like the rabbit hole goes deep.

 

From a legal standpoint, the training side is pretty clear. It's defensible, especially when those works were explicitly put into the public domain. Even if it was not explicitly put into the public domain, there is no way to extract the training material from the model, save for some overfit material that the AI saw repeatedly with the same sample repeatedly.

 

It's theoretically possible to make an AI voice replicate exactly how someone sounds, enough when mixed with something else (eg a backing track, or a laugh track, or some other masking noise)  that it would convince people that the AI is actually the real person. Which is why AI's should be labeled as "AI generated" if they reference the training material, or be given their own unique "name" if they are not intended to have commercial value based on their training source.

 

Eg, if you take a dozen people you like, and train an AI to sound like them, but you give them all cute nicknames instead, that's fine. Until you decide you want to produce content with those voices and people start asking if that voice is NAME-THAT-WAS-TRAINED-ON. It would be both truth and a lie to answer it. Yes that is an impression of "NAME" but it is not "NAME" since it's an AI. People are still not past the idea that it's not illegal to do an impression of someone, be it their voice or their art style, or game mechanics, or any other 'idea' that you can't copyright. What you can claim instead is that the AI trained on NAME, has used "NAME" to step on NAME's brand or trademark. That's where you get in trouble. And that is something that ALREADY exists and requires no additional laws to enforce.

 

You can't sell a POKEMON fan game, It's not a real Nintendo product. We already know this. So making fan-content material of a REAL person, for commercial use, logically, is Illegal. It's just a question of if that REAL person gives enough of a care to go after this infringement of their brand, or if they will look the other way. So instead of a fan game, you just call it something else "POKE-CUTE" or whatever. And that also applies here.

 

But I digress, I think people make mountains out of molehills over AI, and as much as I think there is room for AI to improve, people seem more worried over the possibility that they will not make money from it, or someone else will make money by cutting them out of the process, when nobody actually cares that much to go after small creators. One is not going to be put out of a good job by AI, only one that was crappy to begin with.

 

[HenrySalayne]

2 hours ago, Kisai said:

If I take a commercial recording, regardless of it's source (eg film, music, podcast, talkshow, etc) and use about 1% of it to train an AI. That is "fair use". The actual training of it can always be justified as fair use. It has zero impact on the commercial work.

If you use 1% of 10 million works, you will have a hard time arguing it's fair use. Good luck with that...

 

2 hours ago, Kisai said:

From a legal standpoint, the training side is pretty clear. It's defensible, especially when those works were explicitly put into the public domain. Even if it was not explicitly put into the public domain, there is no way to extract the training material from the model, save for some overfit material that the AI saw repeatedly with the same sample repeatedly.

Public domain? There are a whole lot of different license types and most of them don't allow commercial use and / or require that the author / source be named. Or they are fair use.

And you cannot extract the training material? Refer to the post you quoted...

 

2 hours ago, Kisai said:

But I digress, I think people make mountains out of molehills over AI, and as much as I think there is room for AI to improve, people seem more worried over the possibility that they will not make money from it, or someone else will make money by cutting them out of the process, when nobody actually cares that much to go after small creators. One is not going to be put out of a good job by AI, only one that was crappy to begin with.

People are entitled to their own opinion, but some people's opinion - oof...

 

"You are creating your AI with my works?" - "Yes, but you're a crappy creator and you won't get any money."

 

I know this is a wild and really hard to understand concept, but if you don't want to pay for someone's work to be used in your AI maybe don't fr***** use it? Seems to be the obvious choice if it's sooo crappy, you can't be asked to spend money on it...

[wanderingfool2]

14 hours ago, HenrySalayne said:

Yes, there is.

Why?

 

Simply put the reason why, an impressionist listens to voices/music/whatever their subject is and picks up on the nuances of their voice including tonal, pitch etc.  So a claim would be that the AI is simply processing the data in a similar way (depends how they structured it).

 

If the claim is that AI should not be allowed to train on this type of data, then one could equally claim that humans shouldn't be allowed to.  Yet we still allow impressionists as an impression.  The artist sold the track to be able to be listened to, so all the privacy/personal data stuff gets thrown out the window.

 

As a human, you are allowed to listen to a song and sing your own song in the same style/sound as the original artist so to that extent it should be allowed (now if you don't make it clear that it's not the person though then you could get into trouble as it's starting to intrude using the likeness of another individual).

 

7 hours ago, HenrySalayne said:

You can extract training data from diffusion models. Coincident? No, that's how training works. Just because it's probabilistic doesn't mean you're algorithm creates something "original".

Things like image ones are a whole different beast than voice.

 

For voice, you can get the AI to learn more about pitch/vocals/etc a lot better.  To extract the original would be impossible, it's why you can clone someones voice in now as little as 30 seconds.  A lot of the finer-grain detail work on how people speak has already been built in with other sorts of training data.  What the 30 seconds one does is analyze in specific ways to generate essentially a profile of how someone speaks.

 

Also, in terms of copyright of an output product it wouldn't matter if it had memorized the entire song...the output which is publish is what actually matters in a case like this, as copyright is about how transformative something is.  An example of this would be if you ask an AI to draw an image you know in the style of van-gogh.  Underlying it might still have a painting stored, but it can output an unique image that has your image drawn in the style.