Text and data mining for training data

Bartz v. Anthropic, Northern District of California, June 23, 2025: Judge Alsup held training on purchased books to be fair use, but not the library built from pirated copies. The settlement that followed, $1.5 billion, about $3,000 per book, closed the case over the pirated copies. Training a model means technically reproducing protected works. That falls under copyright, and legal systems answer very differently. EU law created a text and data mining exception with no restriction of purpose. Outside scientific research, though, the rightholder may opt out in an appropriate manner, notably by machine-readable means. The American system has no equivalent exception and reasons through fair use, case by case. Its criteria, in particular transformative character and effect on the market for the work, are precisely what current litigation is testing. The economic stakes go beyond training: an adverse ruling would make unlawful models already deployed in thousands of applications, and there is no technical way to remove the data selectively. For insurance, the exposure sits in intellectual property cover and in the indemnities providers have begun to grant their customers.