Hacker News Reader

Comment by Barrin92 | Hacker News Reader

Don't understand this discussion at all. There's no need to flip your position due to AI. Any produced work may or may not infringe copyright. If you produce a work that lifts say, characters or expressions straight out of a story you're likely violating copyright. If you used AI tools to produce that same work it still does. It's irrelevant how you technically produced the work, by pen, typewriter or chatbot api.

Asking if training a model violates copyright is unintelligible because 100 GB of weights don't resemble any work unless you produce and publish inferences from it.

View on HN

> 100 GB of weights don't resemble any work

I don't think that argument has a lot of weight. By the same logic a zip file of a book doesn't resemble the original work if you look at the raw bytes but you can extract it and get it back. It would be hard to argue that the copyright violation happens only when you extract the zip, and not, say, if you distribute the archive.

LLM could even be seen as a form of lossy compression after all

A sufficiently overfit model is indistinguishable from bad compression.

The same logic doesn't apply because you can't get the data back out of the model by unpacking it. It's theoretically not possible because the model is magnitudes smaller than the totality of the data.

To use, and invert the example from the other commenter, you can even get an existing poem out of a generative model that genuinely was not in the training data at all. This is because it's not an archive, which does correspond directly to one particular work, but a generative technology.

If I create an original file in Photoshop that's 8K, then produce a JPG of it, the JPG is both orders of magnitude smaller, and clearly a rendition of the same work. There's no way to get back to the original 587MB 8K PSD from the 87KB 1024x768 JPG, but that's irrelevant to whether the latter is a version of the former.

Just because the models are not perceptually the same as their training data does not mean that they are not effectively a lossy compression of that data.

But people absolutely have extracted individual source images out of the models by not even particularly careful prompting.

That's absolutely not true.

You might get replicas of incredibly popular works like Mona Lisa due to overfitting but that's it.

If I am wrong please do provide an example as this is very relevant and interesting (and impossible from an information theory pov).

https://www.technologyreview.com/2023/02/03/1067786/ai-model...

The approach I've seen is to prompt for people with unusual names, they're often be only a single source image in the input data set that gets reproduced by the AI.

I've seen examples with the AI "generated" images and the source image side by side - I'll try and find them.

EDIT: Link to the paper https://arxiv.org/abs/2301.13188 and this articles has the example images: https://www.theregister.com/2023/02/06/uh_oh_attackers_can_e... Look at the Ann Graham Lotz image and tell me that isn't the source image being reproduced in a lossy manner.

Right, the LLM is more like the zip algorithm than the zip file. Yes, if you feed certain bits (the zip file) to the zip algorithm, you can get a copyrighted work out, but that means the zip file is violating, not the zip software.

So if I lossily compress a novel so random typos and spelling errors are introduced then it isn't copyright infringement?

Wow, that changes everything. BRB, I got books to pirate.

> Asking if training a model violates copyright is unintelligible because 100 GB of weights don't resemble any work unless you produce and publish inferences from it.

I agree, but there are many who don’t, and assert that any training use should be banned. Those people are the ones who the article is talking about: people who suddenly became copyright maximalists due to their desire to halt AI training.

I don't want to halt AI training, I want corporations to fuck off from using my (A)GPL code to train their proprietary models which they then sell to people writing more proprietary code. I would be ok if the derived code is properly GPL licenced too.

I suspect many people feel in a similar way too (for example, artists whose art is used to train image generators without compensation).

What about people reading your code and learning from it before implementing their own code? What's the similarity level where that becomes a problem for you, if their code is closed-source or uses a license you disagree with?

Putting aside the fact that what we call AI today is not learning in the same way as humans. They operate on a VASTLY different scale compared to humans. On a good week I can read a book. A single book. A massively parallelised data centre can do that billions or trillions of times faster. Scale of effect (lacking a better phrase) must be considered.

a rack of equipment does not need to sleep, eat, take care of themselves, earn a living and so on while churning through millions of words a minute. An actual thinking and learning person has to choose what to spend their limited time and money and attention on, while reading at a pace of dozens of words a minute. Those are not the same things at all.

More than that, AI-generated work is clearly derivative for the simple obvious reason that none of it would exist without the original sources.

I mean, by that logic every fantasy story ever is a derivative work. Should everyone be paying JRR Tolkein's estate royalties the moment they include elves in a story?

That's... not the legal definition of a derivative work.

I agree with this, but is this something that should be dealt with in the law or in the license? My gut feeling is that the just remedy is A) the GenAI models out there should get to do what they want as long as they are not violating licenses, B) the libre software world needs to hustle and release new versions of the appropriate licenses that specifically forbid use of the source code to train AI unless the AI itself is licensed permissively.

Note that in regards to A) I'm pretty sure the AI firms ARE violating copyright today, they have done this knowingly, and they should get a hard slap for that. But they are not violating any particular copyleft licensing provisions to my knowledge

> I'm pretty sure the AI firms ARE violating copyright today, they have done this knowingly, and they should get a hard slap for that.

Depends on the jurisdiction. Do note that many countries already passed laws indicating that training is NOT copyright infringement. The EU[1], for example. In which case, no license would matter.

https://www.reedsmith.com/en/perspectives/ai-in-entertainmen...

[1] Yes, in the EU, you can opt-out (but only for commercial purposes). In other countries such as Singapore however, there are no legal mechanisms for opting out.

> I suspect many people feel in a similar way too (for example, artists whose art is used to train image generators without compensation).

Just to be a counter-voice, I don't. My code is AGPL too, but since the number of copy"righted" things outnumber the number of AGPL things, I'd rather anybody have the ability to train their own AI on all material. Conversely, if it was considered a derivative, only large corporations like Adobe or Microsoft would be able to train on it (e.g. they can just give themselves a license to do so via the ToS).

In other words, it's probably a bad idea to strengthen copyright law for the purpose of enforcing copyleft, due to possibilities of it backfiring on us.

I'm not sure it's always true to say "100 GB of weights don't resemble any work".

If I train a model on a famous poem or something, and it turns out if you ask the model to quote the poem verbatim, it can, then the model contains the poem. Have I not "copied" the poem into the model?

You can simply copy the poem in much less space if that's what you want, but incidental replication is not the same with copying.

When you train a model, you use 1T...20T tokens, with deduplication to reduce direct memorization. Then the model is a superposition of gradients from trillions of tokens, no longer just a copy of a poem, it could also write a commentary about it or compose new ones.

GPT4 summarizes this in a haiku:

"Superposition bright,

Gradients blend in the night,

New poems take flight."

> GPT4 summarizes this in a haiku:

This has never added any value to a comment. Stop doing it.

It’s also not a haiku afaict.

Doesn’t superposition have five syllables?

yep

better than I would have been able to make, anyway

You're selling yourself short. It's not hard to count syllables.

Also note, it's not a haiku

My understanding is that the law ultimately cares about the market impact on the copyright holder, and balancing that with the interests of the public. I don't think the courts will be swayed by arguments based on technicalities like this.

The law doesn't care about the technicalities...? It damned well better.

"Technicality" here means an argument that follows a strict/formal interpretation to arrive at a conclusion that is inconsistent with the principles and objectives of the law.

The law doesn't care about technology technicalities.

It cares about legal technicalities.

The law doesn't care, for instance, that the MP3 file of Beyoncé's latest hit is not the same bits in the same order as the original master file created when she recorded it. That holds true no matter how many times you re-encode it into different formats.

As things stand, there has been no definitive ruling on the matter of ML model training and copyright; I don't know how that will end up shaking out, and honestly, I'm not yet sure how I think things should shake out in any detail. I think that people whose works are being used to fuel massive profit-generating engines for the very wealthy should have some right to give or revoke consent, but I'm not at all sure offhand if there's a realistic way to do that without serious unintended consequences.

But I am sure that arguing that the precise technical nature of how the transformations are performed means that nothing is actually being "copied" should not, and is unlikely to, hold much legal water.

Copyright exists only on complete works and not characters; however any ai produced materials which use trademarked characters would violate trademark. It's just that ai cannot violate copyright unless it produces an identical copy of an existing work.

Characters can definitely enjoy copyright protection, distinct from trademark protection. At least in the US.

Regarding copyright on characters in the United States [1]:

> US Copyright Statute of 1976 does not explicitly mention fictional characters as subject matter of copyright, and their copyrightability is a product of common law. Historically, the Courts granted copyright protection to characters as parts of larger protected work and not as independent creations. They were regarded as ‘components in a copyrighted works’ and eligible for protection as thus.[5]

But in practice, if you write a fanfiction story which uses Mickey Mouse as a character and get sued by Disney for copyright infringement, you will not win in court regardless of how different your fanfiction story is from any Disney story unless you can afford as many lawyers of comparable quality as Disney can. And even then, who knows?

[1] https://en.wikipedia.org/wiki/Copyright_protection_for_ficti...

The binary structure of an mp4 file of a copyrighted movie also doesn't resemble any work, until you interpret it in a specific way.

Someone tried that with MP3's a while ago, as expected they lost: https://arstechnica.com/tech-policy/2009/11/about-those-beat...

> So defiant, in fact, that MRT had actually registered for copyright protection on the songs it was selling. "To make matters worse," says EMI's complaint, "Defendants recently sought to register their infringing sound recordings with the Copyright Office, apparently claiming that because they copied the sound recordings using their own computer system, they now own these digital copies and have the right to distribute them to the public."

Examples of crude MJ plagiarism.

https://twitter.com/kortizart/status/1588915427018559490/pho...

Reply on news.ycombinator.com