HUGE decision on AI and copyright came down.

Training AI on copyrighted material is fair use.

Piracy is considered a separate issue.

Which means that AI companies will be able to make deals with publishers - much like libraries do - to get discounted copies of copyrighted materials for training AI without worry about infringing on copyright because training is considered a transformative and fair use of the material.

Alsup's opinions are always well-written and clear. I wouldn't go celebrating if I were Anthropic (or Open AI, etc). The recitation of facts, and the opinion itself, are actually quite damning of Anthropic's practice of wholesale piracy of copyrighted works, regardless of the use to which it puts it. Here are the short takeaways: 1/x

The act of using a copyrighted work, legitimately obtained, to train the AI on how to use words, isn't the same thing as OUTPUTTING copyrighted words of the author. That wasn't at issue. If Anthropic's AI publishes the content of what the author wrote, or a derivative of it, that could also be illegal (but was not at issue in this motion). 

4/x

Anthropic wasn't doing a lot of that. They were pirating all the books they could find, no different than people who used Napster to download songs. (Remember all the people who got in trouble for that?) Yes, still illegal, even if you later go on to use them to train your AI. And NOT using those pirated works to train AI doesn't excuse HAVING PIRATED WORKS to refer to later or just store. 

3/x

1. IF Anthropic buys a paperback legitimately, tears the binding apart, scans the words into digital, destroys the original, and uses that one digital copy to learn how to write effectively, compute how words go together, extract ideas that can be output in different ways, then YES, the use is transformative and therefore fair use.
 
HOWEVER, 2/x