Anthropic's $1.5 billion copyright settlement reveals their hypocrisy: they train on all creators' output without permission but call it IP theft when Chinese companies do the same to them.
Sax and Chamath argue that Anthropic's position is hypocritical—they believe they can train on everyone's content under fair use, but they call Chinese distillation of their models IP theft, and they have never publicly made that claim because it would poison their own fair use lawsuits. ✦ AI generated
David Sax · All-In Podcast · 2026-07-24 · original ↗
starts at this moment · 50:10
Let's just be very clear about what this judgment was and was not. So, okay, what Anthropic did is they pirated all these books from LibGen and they trained on them. The reason why they got in trouble is because they basically took stolen books. They didn't even pay for one copy of them. But if they had paid for just one copy of each book, they could not have been nailed for piracy. They would have been potentially under fair use, which I understand Jal is still being litigated in the courts, but that would have been their defense. So, the reason why they got nailed with this $1.5 billion judgment is they wouldn't even buy one copy. It is still Anthropic's position and it's OpenAI's position that they should be able to train on all these books under fair use if they buy one copy. And that issue has not been resolved yet. Now you should be able to see the total hypocrisy of their point of view relative to the previous issue which is they believe they should be able to train on every creator's output in the world. You know as long as I guess they bought one copy of it against the will of those creators whether those creators like it or not. They believe it is fair use to train their models and derive their own weights based on fair use. However, they say that the one type of content that you should never be able to train on is their output. That is currently their position. It's completely hypocritical.
verbatim transcript · starts at 50:10
50:04your premonition here Jal. This is like >> I am going to go with for my biggest winner for >> training data owners like the New York Times, Reddit X, Twitter, YouTube etc. I think what we learned in 2023 was that the language models are starting to hit parody very quickly and that the real value is going to be in and it may even become commodities and open source may
50:28win the day. So then I think the winner is folks who have the training data. >> You know the best thing about this is watching Jason's reaction to Jason. Did you do a picture and picture of me just be like go Jac did we make a bet here? Did we make a bet? I don't know. He's gone. >> Well, actually, this I'm this settlement I don't think quite proves exactly what
50:53you want it to prove. Jal, >> go ahead. >> Explain. Can I can I make a nuance here? >> Of course. Of course. >> So, okay. Look, and you know, obviously I'm not a huge fan of anthropics. I think they're potentially destroying the whole ecosystem for their own purposes of regulatory capture. But let's just be very clear about [laughter] what Let's just be very clear about show anytime,
51:13Daria. >> Yeah. Let's just be very clear about what this judgment was and was not. So, okay, what Anthropic did is they pirated all these books from LibGen and they trained on them. And the reason why they got in trouble is cuz they basically took stolen books. They didn't even pay for one copy of them. But if they had paid for just one copy of each book,
51:38they could not have been nailed for piracy. they would have been potentially under fair use, which I understand Jal is still being litigated in the courts, but that would have been their defense. So, the reason why they got nailed with this $ 1.5 billion judgment is they wouldn't even buy one copy. It is still Anthropic's position and it's OpenAI's position that they should be able to train on all these books under fair use
52:01if they buy one copy. And that issue has not been resolved yet. Now you should be able to see the total hypocrisy of their point of view relative to the previous issue which is they believe they should be able to train on every creator's output in the world. You know as long as I guess they bought one copy of it against the will of those creators whether those creators like it or not.
52:25They believe it is fair use to train their models and derive their own weights based on fair use. However, they say that the one type of content that you should never be able to train on is their output. That is currently their position. It's completely hypocritical. And actually, if you go back to anthropics blog post in February where they defined this concept of industrial scale dissolation attacks for the first
52:51time, they coined that expression. And this is, you know, I worry that people in the government policy makers don't understand that this is all part of a anthropic op. No one used the terms distillation and attack together until Anthropic wrote that blog post. Distillation was simply an industry standard practice. But then Anthropic coined this idea of industrial scale dissolation attacks. In any event, if you go to that blog post, search for the