Hacker News Reader

Comment by bee_rider | Hacker News Reader

On the topic of licenses and LLM’s—of course, we have to applaud sourcehut at least trying to not allow all their code to be ingested by some mechanical license violation service. But, it seems like a hard game. Ultimately the job of their site is to serve code, so they can only be so restrictive.

I wonder if anyone has tried going in the opposite direction? Someone like adding to their license: “by training a machine learning algorithm trained on this source code or including data crawled from this site, you agree that your model is free to use by all, will be openly distributed, and any output generated by the model is licensed under open source terms.” (But, ya know, in bulletproof legalese). I guess most of these thieves won’t respect the bit about distributing. But at least if the model leaks or whatever, the open source community can feel free to use it without any moral conflict or legal stress.

View on HN

If you squint at the GPL then you could argue that every LLM is already under it, because it's a viral license and there's almost certainly some GPL code in there somewhere. I'm sure the AI companies would beg to differ though, they want a one-way street where there's zero restrictions on IP going into models, but they can dictate whatever restrictions they like on the resulting model, derived models, and model output.

I hope one of the big proprietary models leaks one day so we get to see OpenAI or Google tie themselves in knots to argue that training on libgen is fine, but distilling a leaked copy of GPT or Gemini warrants death by firing squad.

The hope is that we can out the onus on them to start suing people or whatever, at least. The US legal system is biased toward whoever has the biggest budget of course, but the defense still gets a little bit of advantage as well.

I think the courts were pretty clear, prove damages. I'm not saying I agree in any capacity, but the AI companies went to court, and it appears they've already won.

There have been copyright office rulings saying that ML model output is not copyrightable, so that last part of the suggested license seems a bit strange, since the rulings could preclude it for code at some point.

Also, it remains to be seen whether copyright law will outright allow model training without a license, or if there will be case law to make it fair use in the USA, or if models will be considered derivative works that require a license to prepare, or what other outcome will happen.

How can AI code be added to any kind of open source license, or would it just be that code that isn't covered under the license (since it's effectively public domain?)?

In most cases it would just be added like all other code as if it was proper/allowed until someone brought a civil suit with enough evidence to claim they had no permission to use the code AND it somehow damaged them, which could be quite difficult and prohibitively expensive.

So I would argue in most cases people will get away with it. We must remember that the only person's opinion that matters on what's actually illegal or not is a judge's.

It seems like AI companies are making a major bet on the idea that the output of their models can be licensed and used in products.

What does copyright have to do with it? It's about distribution.

> There have been copyright office rulings saying that ML model output is not copyrightable

Source? If you're referring to Thaler v. Perlmutter, I would argue that is an incorrect assessment.

> The court held that the Copyright Act requires all eligible works to be authored by a human being. Since Dr. Thaler listed the Creativity Machine, a non-human entity, as the sole author, the application was correctly denied. The court did not address the argument that the Constitution requires human authorship, nor did it consider Dr. Thaler's claim that he is the author by virtue of creating and using the Creativity Machine, as this argument was waived before the agency.

If someone had instead argued that they themselves were the author by means of instructing an LLM to generate the output based on a prompt they made up, plus the influence of prior art (the pretrained data), which you could also argue is not only what he probably did anyway (but argued it poorly), but in a way is also how humans make art themselves... that would be a much more interesting decision IMO.

> Someone like adding to their license

I would assume that clause would be unenforceable. They may be able to try to sue for violating the terms of the license, but I'm fairly confident they're not going to get a judge to order them to give their model away for free even if they won. And they would likely still need to show damages in order to win a contract case.

> I guess most of these thieves won’t respect the bit about distributing. But at least if the model leaks or whatever, the open source community can feel free to use it without any moral conflict or legal stress.

The hope is to flip the script. Sure, the company might not fulfill their full obligation under the terms and conditions they agreed to (in the same way that we all agree to these “continuing to use this site means you agree to our terms and conditions,” they are agreeing to the terms and conditions by continuing to scrape the site). But, at least if a model leaks or some pirates software that was generated by one of their LLMs, they can say, well it was open source.

> they can say, well it was open source

I disagree... if those terms are unenforceable, then 1. someone else's model does not become free just because they say it does, and 2. I'm not convinced that is even legal to begin with.

Typically contracts (a license is a contract) have to be fair and mutually beneficial... I don't think a judge would agree that giving a whole model away for free just because you trained on some of their data, is fair, if there's anything even legally wrong with using said data for training in the first place.

You would also need to show a "reasonable purpose" for such a stipulation in the contract. Giving their product away for free as a punishment doesn't sound very reasonable to me, and I don't think a judge would say so either.

The model isn’t given away for free as a punishment, they agreed that the model and everything it produces is free as part of the terms and conditions. If they don’t think it is mutually beneficial, they don’t have to use the site.

If they don’t like the contract, they can try and get it invalidated, but at least it will be a distributed problem for them.

> they agreed that the model and everything it produces is free as part of the terms and conditions

> If they don’t think it is mutually beneficial, they don’t have to use the site.

That's not how contracts work insofar as legal enforcement though. If a judge finds that clause to be unenforceable, then they wouldn't be giving anything up.

> If they don’t like the contract, they can try and get it invalidated, but at least it will be a distributed problem for them.

Reply on news.ycombinator.com