Apparently chardet got Claude to rewrite the entire codebase from LGPL to MIT?

[bob_zim@infosec.exchange]

@scy @Foxboron It is absolutely a violation for the company which built the model to build a model which emits license-restricted code without following the terms of the license. The model doesn’t commit the violation any more than a photocopier does, of course.

The emitted code cannot be copyrighted at all, but if it emitted the code in a way which meets the terms of the license, the code would be covered by the original license.

[douginamug@mastodon.xyz]

@Foxboron the presenters have a live demo: https://malus.sh/

[gooba42@mastodon.social]

@Foxboron Except the output can't be copyrighted and so the result is public domain. It can't even be licensed anymore.

[blogdiva@mastodon.social]

YUP

copyright is for humans, not automata ―hard or soft.

so, ironically, the prompts are copyrightable but not the output.

so anything you want to copyright should not be prompted into a corporate regurgitation machine, including so-called grammar checkers.

@thomasjwebb @Foxboron @scy

[gooba42@mastodon.social]

@Foxboron Went ahead and added an issue since you can't apply an MIT license to public domain LLM output.

[zkat@toot.cat]

@Foxboron that's... not copyrightable, therefore not licensable?

[glyph@mastodon.social]

@Foxboron @joshbressers @scy verbatim isn’t the question here, the question is infringement. is the output here substantially derivative of previous versions of chardet to the point that it could be considered infringing? US copyright precedent is a muddled mess and I think this could implicate at least one unresolved circuit split. I don’t know what the answer will be but I know I wouldn’t want to be standing in the blast radius of that decision

[slightlyoff@toot.cafe]

@Foxboron If you can't copyright it, you can't license the copyright. Interesting times.

[xgranade@wandering.shop]

@Foxboron It looks like this was the PR?

chardet 7.0: ground-up MIT-licensed rewrite by dan-blanchard · Pull Request #322 · chardet/chardet

Python character encoding detector. Contribute to chardet/chardet development by creating an account on GitHub.

GitHub (github.com)

Even aside from the ethical and moral issues with LLMs, it doesn't seem optimal that a 15k line PR affecting almost a million dependent repos (if GitHub's count is to be believed) was up for three days before getting merged in.

[xgranade@wandering.shop]

@aud It's at least not systems code, so there's not a lot of potential for buffer overflow and other memory unsafety exploits, but yeah. No. chardet is not a small surface area.

[wordshaper@weatherishappening.network]

@thomasjwebb @Foxboron @scy In the US, at least, human authorship is required for copyright, and if you try to copyright something that's a mix of AI and human generated then generally only the human generated part is copyrightable.

https://www.congress.gov/crs-product/LSB10922#:~:text=Granting%20that%20human%20authors%20may,applying%20to%20register%20their%20copyright.

This is separate from the LLMs emitting text other people have written, so at *best* this code can't be licensed because it's not copyrightable, and at worst its license laundering and there's precedent (IIRC) for stomping on that hard.

[aud@fire.asta.lgbt]

@xgranade@wandering.shop There's just no way that's a good idea. I'm pretty sure a human who tried to push a 15K rewrite into most libraries would be yelled at forever and the PR rejected, or asked to be broken into smaller PRs, because it's just such a large change in one go and no one can possibly fit that entire thing into their head.

It doesn't magically become a good idea just because claude shat it out.

[jens@social.finkhaeuser.de]

@scy @Foxboron It's a bit complicated, actually. IANAL, but this is what I understand:

- The music notation is copyrightable, individual notes are not. A sequence of notes is debatable, and it depends highly on recognizability AFAIK.

- A music recording is copyrightable. Playing that music in a distinctly different arrangement, less of an issue.

- Arguably, a change in digital format is either still the same recording, or sufficiently indistinguishable from it.

- Copyright has an ancient...

[jens@social.finkhaeuser.de]

@scy @Foxboron ... naming and goes back to a time where making copies and distributing them was the hard part.

This is a non-problem in the digital age, which is why it's fine to create backup copies of copyrighted works, so long as the people accessing them are always the people having purchased/licensed an original copy.

So LLMs training on GPL is not itself a copyright violation, and them reproducing similar code isn't either, but then publishing such sufficiently similar code is.

[jens@social.finkhaeuser.de]

@scy @Foxboron TL;DR what others already wrote: if the result is similar enough to inputs, the copyright holder of the inputs could challenge it, yes.

[jens@social.finkhaeuser.de]

@scy @Foxboron If courts decide to throw this out, I would personally *love* for someone to use the exact same argument to produce a minimally altered copy of Avatar, and have Hollywood throw a fit.

[jens@social.finkhaeuser.de]

@scy @Foxboron Basically, let's not fight this, let the industry giants fight each other. Throw a few near-copies of Metallica songs in for good measure, so we get a v2 of that "Napster baaaad" animation with greedy gnome Lars Ulrich.

[jens@social.finkhaeuser.de]

@scy @Foxboron Either LLMs will die on the spot, or Copyright does.

[mutesplash@uncontrollablegas.com]

@Foxboron Fun, one of the fundamental problems I have with this technology!

[foxboron@chaos.social]

@xgranade
They have been the upstream maintainer for years, so I don't see any huge issue with that.

I would have done the same probably?