@florian It would, if and only if:

1) You train only on compatible license content
2) You meet all the attribution requirements of those licenses (this is a big one)
3) Your weights are licensed under a compatible license themselves (usually the same for most copyleft licenses)
4) You understand that the output of the model may be copyrighted and require the same licensing and attribution (making the model unsuitable for, say, creative generation for publishing, but it would still be fine for a local voice assistant or something like that) and make your users aware of this.

Unfortunately, I haven't been able to find any models that meet any of those conditions, let alone all of them (other than KL3M which credibly claims to be trained on PD and non copyrightable content only).

@florian I don't think there's any need to care about where a work is encoded. You do need a list of all authors though.

I think there's some flexibility in that you might not need to literally list every individual, though IANAL. For example, if you train on Wikipedia (only), you could plausibly get away with specifying the exact database dump you used (where the edit history data is available) without having to extract and collate the author list yourself. But if you're scraping something like GitHub that does not make explicit dumps available, yeah, you'd better at *least* list every project and commit ID you scraped (and even that might not be enough).

@lina @florian

Who has standing and losses to legally attack anyone for this?

Say I didn't follow these (arbitrary...) demands and I downloaded a MIT project from github as anyone can, and I put it into an LLM.

The maintainer of the MIT project has enough mental flexibility to look up from the concise, liberal text of the MIT license, and formulate a complaint to the judge about his losses and ownership of the (completely different) code the LLM emitted?

I guess we will all see, right?

@hopeless @florian Your argument is basically "open software licenses don't matter because nobody is actually going to sue people for violating them"

This is not a good argument.

@lina @florian

My point is that to attack someone through the courts, you must have standing - it's your copyright - and be able to show damages.

The people who licensed their work under MIT are already at peace with $0

"Permission is hereby granted, free of charge, to any person obtaining a copy of this software... to deal in the Software without restriction, including without limitation the rights .."

What damages would they show? Copyright maximalism is not compatible with FOSS.

@hopeless @florian That's really not how it works.

https://www.law.cornell.edu/uscode/text/17/504

They get to claim the profits made off of their copyrighted work, or statutory damages, at their choice. The latter doesn't require any money to move anywhere. You can be on the hook for $30k, or $150k for willful infringement.

@lina @florian

No it is exactly how it works.

> "... the copyright owner’s actual damages and any additional profits of the infringer..."

Actual damages for works under MIT: $0. Additional profits when your own work is FOSS: $0.

> "... statutory damages..."

"...to be eligible for statutory damages and attorneys_fees, the creator must have registered their work with the u.s._copyright_office before the infringement occurred (or within three months of publication)."

https://uslawexplained.com/17_usc_504

@hopeless @florian Right, except:

In establishing the infringer’s profits, the copyright owner is required to present proof only of the infringer’s gross revenue, and the infringer is required to prove his or her deductible expenses and the elements of profit attributable to factors other than the copyrighted work.

So they get to claim all your revenue and you are on the hook for proving none of it derived from the infringement.

That is not easy. For example, if you only released the model for free online, but then landed an AI job, you would have to prove the model release was not a factor in you landing that job.