#144. ♟️ Anthropic's $1.5B settlement was 4D chess at its finest

This is educational material and does not constitute legal advice nor is any attorney/client relationship created with this article, hence you should contact and engage an attorney if you have any legal questions. No warranties, express or implied, are made with respect to its accuracy. Information contained herein, or information relied upon, is subject to change without notice.

On September 5, Anthropic got slammed with the largest U.S. copyright settlement in history: an unimaginable $1.5B based upon $3,000 per pirated book that Anthropic scraped for training its language models.

Crucially, however, the court didn’t take issue with the training itself, an “exceedingly transformative” process, and “spectacularly so” at that — largely reflecting my own research — but rather that Anthropic chose to store the pirated works in a massive digital library for future use.

For the record, I'm a huge fan of Anthropic, I love their Constitutional AI, I think Claude Code is magical, etc.

I also think that their decision to build and keep that central library of copyrighted books — many sourced from shadow libraries — was so (seemingly) irrational (but see below) on its face that no competent legal or exec team could have ever authorized it.

Unless it was worth it. Hence the only logical conclusion is that it was intentional.

Hear me out:

Anthropic just announced a $13B Series F at a $183B valuation. Like other frontier labs, they’re burning billions per year. So while $1.5B wasn’t trivial, it wasn’t catastrophic.

In fact, one could say it was just the price of doing business: they budgeted that training on books and then warehousing the entire corpus would be worth at least $1.5B, hence why they did it and settled accordingly. (Frankly, they probably budgeted even higher; good timing, that Series F, which coincidentally landed the same week that the court dropped its decision!)

My point is, I don’t think this was just a careless fumble or a legal lapse; I think it was a brilliant game of 4D chess.

For some reason, copying > training > and then discarding the copyrighted material, as would have been proper to avoid liability, wasn't sufficient for Anthropic; they had to store the copyrighted material indefinitely, even if it cost them $1.5B.

The question is ... why? What value does that entire corpus of copyrighted material hold ex post facto, i.e., after the training is complete (fine-tuning, debugging, future development notwithstanding)? Why didn't they just simply discard it like everyone else?

So to paraphrase that infamous Anakin/Obi-Wan exchange, Anthropic never asked for a $1.5B bill, but it’s what they wanted.

I find this as puzzling as it is fascinating....

Looking for past newsletters? You can find them all here.