I’ve been wondering: what’s stopping the music industry from creating a new, legally encumbered audio format that explicitly prohibits use in AI training?

The format license could include terms of service and an arbitration clause, with strong penalties for misuse. For example: no reverse engineering (which would also technically circumvent of access controls on copyrighted works, a violation of the DMCA), and no use of the format’s contents (i.e. encoded music) for machine learning purposes.

As a further deterrent, the license could include a "poison pill" clause—if violated, any AI models trained on the format must be released for unrestricted public use.

It wouldn’t stop bad actors entirely but put a crimp on things. If music is only released in that format AI companies could never claim that they legally acquired the music to train on.

sr registrations already give owners sole discretion to say yay/nay as well as are the defacto ownership proof when plaintiffs sue based on infringement of source....as is now in play in existing ai court cases of the past 18 months.

Sr protection is in place as a global standard, recognized worldwide in court cases, and has been the great protector since copyright registration began in the previous centuries.

Most of my wealth was made because of my sr registrations. It's a very ironclad type of protection.

No need for a further invention...much less an "audio format"..... do you want to pay to invent one?

The AI companies disagree. Japan disagrees. They probably have already trained on your works.

The biggest task of such a minor DRM protection would be adoption, not making the protocol. It doesn't have to be uncrackable, it just has to be a business barrier to AI companies.

Hi there, I'm sure you made 5 mistakes in spelling word china. :D

You must be young.

All drm can be hacked....all supposedly poison pill types of file tools introduced in 1997-ish constantly failed...no one wants to pay to create another hackable format.

If Xyz in Japan trains via my sr regs, and then releases a product that is lifted (audio fingerprints being the key to an sr prima facia source)...and I find out....I sue in New York or here in LA.

Defendants then are required in subsequent filings and trial to disclose...

(1)..if it wasn't "my" registered audio sourced....what DID they use?

(2) Explain and show the supposed "not my source' routine used to create disputed title/product.


Guess how that response has been going in existing cases so far (not my sr regs mind you..I haven't sued an ai company....yet )

Either answer given by defendants...I win.
They sourced "something" and whatever it was, I only file suit when I know FOR SURE that it was mine (the easy part).

And while the penalties are astronomical in civil court, the defendant xyz will likely do an intense scramble to settle out of court for an almost equal astronomical amount.

It's simply not in the biz interest of company xyz to be inviting lawsuits (unless they're incredibly stupid...which some are).

It's much more conducive for xyz to ask for a license to use the sr.....which...although for possibly an astronomical price....I (might ...if me) say yes.....or more importantly....it's my right to say "no".

The existing rules are fine.

And you know what? The often-stupidity of cutting edge tech companies is they move so fast that they f*** up by not playing by existing rules....and then are sued....by people like me....rights holders....who then ourselves add to our wealth when the f***-ups settle out of court.

A win for rights holders every time (except during the Napster debacle )

Existing statute rules are purrrrrfect.

It's unclear to me why you'd think you'll "know FOR SURE" with AI-generated music.

I think it is impossible to know "FOR SURE" anything about AI, because AI simply mixes millions of snapshots into one.

I could be wrong I'm not a lawyer, but I'm an IT guy, so this is my guess,
he wrote about "audio fingerprints", so most likely his "for sure" ー is about an AI error or random chance, when the final sound in some period of time (regulated by law), match a fingerprints identical to the fingerprints of one of the works, protected by a specific author.
so, if it just matches - he wins the case. regardless of the circumstances.

but here another side of it - AI is really a random generator from billions of parameters, so it can create a similar print in a short period of time, according to probability theory, even if AI did not actually use this work for learning.
but since this is such a huge random - it is impossible to prove that they "clean", so no solution for this, so this pretty much as patent trolling. although in this case I have nothing against it.

I specifically state Japan because it has made it legal to train on your works. Good luck winning there. But if the music a work was distributed solely via the means I propose with contractual limits we're no longer counting on just IP but instead on adding additional contractual protections.

The DRM is not what is to stop the training. Of course DRM can be hacked. But if music was ONLY made available via the DRM'd method LEGALLY companies are bound by the restriction and have no defense legally to no following the restriction. The 'time shifting' and 'fair use' loopholes won't apply if they use them to build a commercial product.

Like you are saying, it's pure lawfair and winnable. Your response agrees more with my initial question than you think.

That all makes sense, but if Japan has already made it legal, how enforceable is it gonna be in the rest of the world? Could a company just set up a nominal HQ there and train their models with impunity?

I bet this alone makes it unenforceable. If the AI trains 90% on your song, and just adds in 10% from some other song, it will be more different from you than Greta Van Fleet is from Led Zeppelin.

Japan added the allowed use to IP law. In my proposal, if music is released with an encumbranced codec, it can only be used legally by following that encumbrance. Think open source. Someone can't later claim they don't have to follow the license to the source code if you can only get it as open source. If you could only obtain a song via agreement to follow the 'no AI training' emcomberance, which the codec automatically attaches via it's use agreement, you are now legally encumbered under that requirement. It moves from IP law to contract law. And AI companies are going to have a hard time stripping contract law (unless they win at pitting every country against each other in a music AI war as a national security need).

It makes IP unenforceable. But AI companies would still have to respect the encumbrance I propose because it's not IP law, it's contract law. And it's on the training side, not the output.

I don't understand how changing which law they are in violation of makes it any more 'enforceable'. Do you think they can't/won't get around your puny encumbrances? How about if they play your 'new codec' out of a speaker and mic it up? After all, what the AI is learning is the content, not the sound quality. I doubt it will be bothered by a bit of generation loss.

No matter how you propose this "works", IMO, the history of such attempts is far from encouraging.

At any rate, if I was making an AI, the last thing in the world I would want to do would be to train it exclusively on a single artist. I would train it on thousands of artists and extract the things they might have in common. It would be like that Arthur C. Clarke story.

Theoretically, you could release the source code of the decoder under a license that restricts the use under certain circumstances (similar to the GPL that requires you to also provide the source code of an derived code).
This wouldn’t prevent anyone from writing their own decoder that can read the files, reverse engineering them. For this, you would need to have a patent on the codec or use some form of DRM. Both are not desirable, imho. Patents expire anyway, and DRM is just restricting the buyers/users.
Also, it would be pretty much impossible in practice, at least on a bigger commercial level, as the streaming services would never agree that anyone provides them with files that deny them using those for training their own models.
(And by the way, the result of those is not necessarily always bad, e.g. stem separation, a tool that many wouldn’t want to do without any more in music production, wouldn’t have been possible without training on millions of songs).

Watch Benn Jordan's YouTube video on a process he has helped develop that encodes a mix so A.I. can't analyze it.

I think its pointless since, using a simple logic, at the end the music should be played by everyone, so why AI wouldnt integrate that decryption codec to its alogrithm? Like reading inside a zip file, if everyone can get into it AI would too by simply adding that layer of decription.
Unless you have a password for it but it would make the music restricted and not to everyone, and no musician wants to do that.

Beside, this fear about not letting the AI read your stuff is over estimated. Our additions to music now it is negligble. Millons of hours of music already recorded and our song is like a drop of water in the ocean. And... is not that we discovered new scales or anything not dont before.

Well Harry Partch did it, but I doubt any AIs will be stealing his stuff.

It seems to me that an AI that is trained on "everything" is a lot like a painter who has does Master studies. The difference being that the human will presumably synthesize his influences and add his own contribution, whereas the AI will never have the "own contribution" part - just the influences. But the point is, that in the end, there could very likely be nothing specific that Franz Hals can point to to say "hey you got that from me"

I bet you could train an AI on 21 Pilots music and everyone would think you trained it on Pearl Jam. But this is probably not the direction it will go. It will probably end up with people training their AIs on everything they can find that is 'catchy', 'hooky' and popular.

Its interesting and philosophical, that you could wonder "what is a contribution"?
Sometimes fusioning 2 styles it is the contribution itself. What was the contribution of Beatles... scales? Not.
New sounds? Yes. New chords, not.
Mixing pop with classical music like In My Life? Yes in many cases. Of course they are the GOAT, but maybe their greatest contribution was fusioning styles.. and that can be done by AI i guess.

Prompt "write and perform a song driven by Piano Moonlight Sonata arpeggio, same key and tempo but 3 voice sing " because the sky is blue..."

Speak for yourself.