>If I can reproduce the entirety of most books off the top of my head and sell that to people as a service, it's a copyright violation. If AI does it, it's fair use.
Assuming you're referring to Bartz v. Anthropic, that is explicitly not what the ruling said, in fact it's almost the inverse. The judge said that output from an AI model which is a straight up reproduction of copyrighted material would likely be an explicit violation of copyright. This is on page 12/32 of the judgement[1].
But the vast majority of output from an LLM like Claude is not a word for word reproduction; it's a transformative use of the original work. In fact, the authors bringing the suit didn't even claim that it had reproduced their work. From page 7, "Authors do not allege that any infringing copy of their works was or would ever be provided to users by the Claude service." That's because Anthropic is already explicitly filtering out results that might contain copyrighted material. (I've run into this myself while trying to translate foreign language song lyrics to English. Claude will simply refuse to do this)[2]
They should still have to pay damages for possessing the copyrighted material. That's possession, which courts have found is copyright violation. Remember all the 12 year olds who got their parents sued back in the 2000s? They had unauthorized copies.
I don't know what exactly you're referring to here. The model itself is not a copy, you can't find the copyrighted material in the weights. Even if you could, you're allowed under existing case law to make copies of a work for personal use if the copies have a different character and as long as you don't yourself share the new copies. Take the Sony Betamax case, which found that it was legal and a transformative use of copyrighted material to create a copy of a publicly aired broadcast onto a recording medium like VHS and Betamax for the purposes of time-shifting one's consumption.
Now, Anthropic was found to have pirated copyrighted work when they downloaded and trained Claude on the LibGen library. And they will likely pay substantial damages for this. So on those grounds, they're as screwed as the 12 year olds and their parents. The trial to determine damages hasn't happened yet though.
Assuming you're referring to Bartz v. Anthropic, that is explicitly not what the ruling said, in fact it's almost the inverse. The judge said that output from an AI model which is a straight up reproduction of copyrighted material would likely be an explicit violation of copyright. This is on page 12/32 of the judgement[1].
But the vast majority of output from an LLM like Claude is not a word for word reproduction; it's a transformative use of the original work. In fact, the authors bringing the suit didn't even claim that it had reproduced their work. From page 7, "Authors do not allege that any infringing copy of their works was or would ever be provided to users by the Claude service." That's because Anthropic is already explicitly filtering out results that might contain copyrighted material. (I've run into this myself while trying to translate foreign language song lyrics to English. Claude will simply refuse to do this)[2]
[1] https://www.courtlistener.com/docket/69058235/231/bartz-v-an...
[2] https://claude.ai/share/d0586248-8d00-4d50-8e45-f9c5ef09ec81