Paying AIs to Read My Books (kevinkelly.substack.com)

Anthropic agreed to a roughly $1.5B settlement in Bartz et al. v. Anthropic after being accused of copying a 500,000‑book “shadow library”; the court awarded roughly $3,000 per affected book and published a searchable list of titles. Anthropic says it didn’t use that specific library to train released models, but the ruling penalized unauthorized possession of copyrighted copies regardless of whether they were used in training. Authors may receive only part of awards (publishers often take large shares), and class‑action proceeds can be consumed by fees — so practical compensation may be limited. The case underscores a deeper shift: the primary audience for many non‑fiction works is becoming AI systems, not just human readers. Because LLMs reflect and can be steered by their training corpora and fine‑tuning (including RLHF and targeted datasets), inclusion in training data confers outsized influence on how ideas circulate. That creates incentives for authors or publishers to pay for inclusion, collaborate on “AI‑friendly” formatting, or lobby for new licensing constructs (e.g., a Right of Reference) rather than rely on copyright’s copy‑protection model. It also raises equity issues—languages and obscure works may be excluded—and signals that curation of training corpora will become both a technical and commercial lever shaping what AIs know and recommend.