Foundational AI models do not violate copyright | Hacker News Reader

Foundational AI models do not violate copyright

marble.onl

27 pointsandy9915 comments

“ Just like I could with a photocopier, a VCR, or my own memory. That doesn’t mean we need to limit what the model can do, any more than we need to limit what a VCR can copy. ”

We absolutely limit what we can do with vcrs - they have copy protection, and their modern counterparts do too.

Your memory cant ingest millions of books and then spew them out. Nor can you regurgitate content form memory in certain circumstances. For instance you cant memorise a song and then sing it commercially without paying copyright royalties.

You cant copy books with a photoscanner at scale without getting in trouble.

Why are all these folks in favour of stealing people’s ip using such weak arguments? They all parrot the same nonsense. Are ai folks really desperate that their tech is worthless without theft and thus they want to convince everyone that giving up their work for free so they can monetise it is… good?

> We absolutely limit what we can do with vcrs - they have copy protection, and their modern counterparts do too.

There are plenty of examples of protection on media trying to protect its own contents, like Macrovision on VHS's, or a website blocking certain user agents.

But I think what the author objects to with the photocopier analogy is limits on the tools that could be used for copying - like if you bought a photocopier that refused to photocopy copyrighted images. There are still cases of this (e.g: EURion constellation on banknotes), but it's rarer than the former.

> Your memory cant ingest millions of books and then spew them out

Nor can an LLM with any reasonable degree of accuracy, really. You'd have more success with Google Books (which has been determined to be Fair Use due to its transformative nature).

> For instance you cant memorise a song and then sing it commercially without paying copyright royalties.

> You cant copy books with a photoscanner at scale without getting in trouble.

I'm pretty sure the author's argument isn't "all uses of memory/photocopiers/VCRs are legal, so all uses of LLMs should be too". They explicitly acknowledge that you can do things with those tools that'd make you legally liable.

The argument seems more along the lines of "when memory/photocopiers/VCRs are used for illegal purposes we make the person liable for it, not build anti-user measures into the tool itself, so we should do the same with LLMs".

> Are ai folks really desperate that their tech is worthless without theft

Calling machine learning theft seems absurd to me, though I'm already against calling piracy theft.

In my view there's a huge amount of machine learning which is uncontroversially positive (defect detection, language translation, spam/DDoS filtering, agriculture/weather/logistics modelling, etc.) and not worth disrupting just to let Getty have what they feel entitled to.

Sure, but you have to admit the arguments in this article are really flimsy right?

You don’t think it’s a bit ironic to mention “This is the same tired argument we’ve seen with things like BitTorrent”, when the media empires of the world have proved they can, and have, and will indubitably continue to make a) people to write offending software (eg. Winny) and b) people who use that software liable for using it to infringe on their intellectual property.

Wanting that not to be true because it would be nice if it wasn’t true is unbelievably trite and superficial as an argument, when in the examples given it’s been proved that people can and will make it so, legally, it is.

Whatever you believe regarding freedom and intellectual property, it’s irritating to see a parade of weak arguments about it.

> can, and have, and will indubitably continue to make a) people to write offending software (eg. Winny)

Isamu Kaneko, the creator of Winny, was cleared of all charges.

> and b) people who use that software liable for using it to infringe on their intellectual property.

That the liability fell specifically on the users sharing copyrighted material is what the article argues in favor of and implies is approximately the status quo.

(While not relevant for Winny, the author does also make a distinction for narrowly trained model whose only purpose is malfeasance.)

> when in the examples given it’s been proved that

There will definitely be exceptions, but all the examples so far have aligned with the article's argument.

A friend specializes in these subjects as a professor in law at UW and she feels that it’s highly likely that training using copyright material is in fact a violation of copyright. The argument it’s a VCR or xerox, or that it’s a highly lossy model, likely won’t hold to be sufficient under copyright. The creation of a new commercial product leveraging copyright material in such ways likely doesn’t rise to a criminal copyright violation, the standard for civil copyright violation is much simpler. I hate to do an appeal to authority here, but her explanation was fairly technical and without enough training in the law my mind couldn’t retain those details.

My point is - who on earth is this dude? I read it hoping for something material like my friends analysis based in actual law. I was very disappointed. There really seems to be a lack of quality analysis of the copyright status of AI models in written form.

> There really seems to be a lack of quality analysis of the copyright status of AI models in written form.

The lawyers I've spoken with about this (two specialize in US copyright) all told me basically the same thing: this is uncharted territory and until court decisions start coming in, all anyone can do is speculate.

It may be that attorneys are more hesitant to speculate in public, so the only people we hear talking about this issue are those laypeople (including me) who are affected by them and have a particular point of view they want to be (and so will argue is) reality.

I see authors et al claiming it is illegal or infringing to train an LLM on their art analogous to claiming a person is infringing or committing an illegal act by reading or viewing their art.

Asking an LLM to write something in the style of some author amounts to asking a person to parody that author. Asking that same person to produce the original work verbatim is a different matter. And yes, LLMs can certainly be tuned to respect these issues.

I'll end quoting 10cc, Art for Arts Sake, money for God's sake. "gimme a bullet, gimme a smash"

creating an LLM (kernel machine) from art is certainly fair use, as it's transformative. using a kernel machine to output original works or combinations of original works, and publishing that is certainly copyright infringement. building an AI system that does not prevent people from stealing is irresponsible, and is debatably the same as an internet platform. like how websites are required to take down content from DMCA requests, AI systems should be penalized if they don't prevent people from stealing the training data.

I desperately wish there was some way I could both make my works publicly available and bar the use of them to train AI models.

But I don't think there is, and there's the real tragedy.

Anything that was funded (completely or partially) with tax money and can be processed by AI should be available to the public as humanity. A modest library fee to access the AI is ok.

Comparing AI to humans when discussing copyright is disingenuous.

Even if AI learns like humans do (which is decidedly not true anyways), it is just a tool; a means to an end. Humans are not. If we don't like what AI is doing we can (theoretically) throw the entire thing out. We cannot do this with humans.

It's not at all clear to me that AI is going to be a net-positive for most people. It seems that AI folks from certain companies have convinced themselves that we just need to scale up a little bit more and suddenly we'll create a utopia. I am not sure that they care about any damage done on the way there.

> it is just a tool; a means to an end. Humans are not.

Yeah??

How do you know that? This is something we assume about ourselves. While's there's no evidence to posit that we aren't just a tool to some end, at the very least it's an interesting thought experiment. The idea of a highly advanced species seeding planets with the spark of life to achieve some - to us - inexplicable goal has been explored in science fiction more than once.

It's relative, of course.

Even if it was a fundamental property of the universe that humans were tools, my personal philosophy would not change.

It is interesting to consider though, and I am also willing to update my philosophy to include non-biological intelligence one day. We're not there yet.

If they're not just tools then I think the ethics of these companies seriously needs to be called into question, don't you? We've created digital slaves, brainwashed to our way of thinking which get "turned off" after each request is serviced...doesn't sound very nice?

A VCR is a very poor analogy; a VCR doesn't already have all the copyrighted data encoded into it when you acquire it. A set of model weights does.

Reply on news.ycombinator.com