AI training data: US court cases reveal more than European transparency measures
The most striking example is that of Anthropic. Its SDS for Claude Opus 5 refers in particular to Common Crawl, specialised datasets, licence agreements, private data and physical books acquired by the company – euphemistically referred to as ‘ acquired physical texts ’ – without ever providing any details about the works in question. However, in the United States, the Bartz v. Anthropic established that Anthropic… had downloaded several million books from LibGen and PiLiMi, two pirate libraries, but had also purchased millions of printed books, which were then cut up and digitised to form training data. The judgement describes the acquisition methods, bibliographic catalogues and the composition of the corpora used by various models. The models and time periods in question are not necessarily identical, but the difference in…