Introduction:
The recent copyright settlement in Bartz v. Anthropic (hereinafter, Anthropic) is now the centre of attention.[1] Anthropic has agreed to pay $1.5 billion for its unauthorised activities, including training a large language model (Claude).[2] As part of the settlement, right holders have the following options: submit a claim (by March 23, 2026), b) opt out (by January 7, 2026), c) object to the settlement (by January 7, 2026), or d) do nothing.[3] Before this settlement, Judge Alsup, Northern District of California, United States, confirmed that training-specific LLMs are ‘quintessentially transformative’ and constitute fair use. However, the court had not exempted the use of pirated sources. This piece explores the legal treatment of using ‘pirated sources’ for AI training under fair use assessments. In doing so, it contrasts the same under Text and Data Mining (TDM) exceptions in the EU and UK law.
[1] Bartz v. Anthropic PBC, No. 24-cv-05417 (N.D. Cal. June 23, 2025)
[2] Available at <https://www.bbc.co.uk/news/articles/c5y4jpg922qo> accessed on 5 September 2025
[3] ibid.
