Bartz v. Anthropic settled at $1.5B; training data provenance is a deal blind spot

alex_verem · x · 2026-08-17

Citing @mardehaym, the author notes Bartz v. Anthropic settled at $1.5 billion over 500,000 pirated works — and zero analysts in the acquirer's data room had flagged training data provenance as a risk before close.

He shares a case he worked on: an acquired company had scraped public medical journals to train a classification model, and two publishers sent cease-and-desist letters within weeks of the acquisition announcement. The publishers identified 200 journal articles in the training set, with licensing fees of $4,000 per article per year over three years of use, and the acquirer's legal team spent four months on it. Lesson: AI deal diligence must cover training data provenance.

Original post →

More from Venture

Venture channel →