Training on copyright was ruled to be transformative.
Anthropic was just fined for not paying for the pirated books they copied into a training database. Same as anyone else who copied pirated IP onto their hard drive.
However, training an AI on copyright has been ruled to be sufficiently transformative and not a violation of IP laws. The same way you can make a gameplay clone of Call of Duty without any issue.
So they got fined for… stealing?
Also, if I clone Call of Duty, and use their skins, make a similar soundtrack, call my maps the same, then there will definitely be an issue.
I support the use of AI and all, but to hide behind transformative use of copyrighted intellectual property is a discredit to the colossal amount of human work and knowledge that these companies pirated and had their models trained on.
>So they got fined for… stealing?
Correct, 2001 napster style pirating of content. You cannot copy stuff you didn't pay for onto your hard drive.
>Also, if I clone Call of Duty, and use their skins, make a similar soundtrack, call my maps the same, then there will definitely be an issue.
A gameplay clone as I stated, tons exist (team death match, capture the flag, battle royale, with first person gunplay and army guys shooting each other). The rest is your argument, not mine.
I don’t think I’d equate a corporation downloading TBs of copyrighted works (with the intent of making a profit on stealing copyrighted works) and John downloading a few GBs of MP3s for personal consumption.
Probably why John wasn't ordered to pay $1.5B.
yep, theft but not plagiarism.
Theft would give the victim a one-time payment, plagiarism would have given the victims payments every time the model was used. In other words: IP stealing through training LLMs is declared 100% legal.
The question is what happens when somebody mass-"uncopyrights" entire libraries of content for the first time. Because, of course, this judgement should have sent the stock price of Disney, Paramount, Warner Bros, New York Times ... to zero.
Why? Because at this point these companies are living on borrowed time. Courts have outlined a clear, legal, process to "uncopyright" any given work.
If I make Bambi 2 and toss it up on YouTube I’m going to be sued by Disney for a wide variety of intellectual property violations, it doesn’t matter if I make it by hand or via an LLM. Nothing has changed for big organizations with a legal team.
Sued ... of course. But what will they claim? The courts have declared that training an LLM on copyrighted data and then having it produce anything DOES NOT make a derivative for the purposes of copyright. That it is fully owned by the person who used the LLM and zero licenses are needed to sell it, rent it out, or whatever.
By that same argument, distilling a model to train your own is also transformative. We can’t have it both ways. Either both training on copyright and distillation are transformative, or both are illegal. Also, it is not a matter of settled law that training on copyright is transformative. That was a single case, and others are still pending AFAIK.
> Same as anyone else who copied pirated IP onto their hard drive
Not really the same. Private citizens have received fines MUCH higher per-work when downloading for just their private consumption.
Ruled by whom? The International Court of the Northern District of California?
What if use an LLM and attach CoD screenshots, or assets?
>"Anthropic was just fined for not paying for the pirated books they copied into a training database."
Normal person would be sitting in jail for doing same thing. It is a shame they call it justice.