How Copyright Lawsuits Shape Model Behavior

aakashgupta · x · 2026-07-18

The author points out that Anthropic's $1.5 billion payment last year for pirated books in training data marks one of the largest copyright settlements in US history. Conversely, if Chinese labs use the same "shadow libraries," they have few reachable assets in US courts, leading to entirely different practical constraints.

They further note that jurisdictions shape US model behavior far more than people realize: facing potential massive statutory damages, labs are forced to choose between settling, filtering, or refusing to train. The problem is that users often perceive these filters as the "model getting dumber"—even if benchmark scores suggest it's actually stronger.

Original post →

More from AGI Musings

AGI Musings channel →