anthropic / free software foundation / gfdl

The Free Software Foundation (FSF) said it may take action against Anthropic over alleged copyright infringement tied to the class action Bartz v. Anthropic, according to a settlement notice it received. The notice says a district court found training large language models (LLMs) on books to be fair use, but left for trial whether downloading those books—sourced from Library Genesis and Pirate Library Mirror datasets—was legal. The parties are reportedly pursuing a settlement and contacting potential rightsholders to offer payments instead of damages. FSF says its book “Free as in freedom,” co-published with O’Reilly under the GNU Free Documentation License, appeared in Anthropic training inputs. FSF argues the proper remedy is “freedom”: releasing complete training data, model, configuration, and source code to users.

The Free Software Foundation says Anthropic included the FSF-published book Free as in Freedom in its LLM training data and warns that using GNU FDL-licensed text without honoring its terms could mean the FSF will seek remedies if its copyright is infringed. The FSF argues that the GNU Free Documentation License requires downstream recipients to receive the full source, training data, model weights, training configuration and software — effectively demanding open distribution of models trained on such material. The notice frames any lawsuit participation (e.g., Bartz v. Anthropic) as likely to seek “user freedom” as a remedy rather than monetary damages, pressing LLM builders to disclose training inputs and model artifacts.

The Free Software Foundation (FSF) has demanded that Anthropic and other LLM developers respect free licenses after discovering Sam Williams’s Free as in Freedom—released under the GNU Free Documentation License—appeared in datasets used to train Anthropic’s models. The FSF argues that freely-licensed works permit unrestricted use and therefore LLMs trained on such content should be released “in freedom”: sharing full training inputs, model weights, training configurations, and source code with users. The demand follows the Bartz v. Anthropic class-action settlement, in which Anthropic set up a $1.5 billion fund to compensate authors for unlicensed training data. The FSF acknowledges it lacks resources for extended litigation but says it would seek user freedom as remedy if it pursued a case. The call spotlights tensions between open licensing and commercial LLM practices.

The Free Software Foundation (FSF) says Anthropic’s LLM training data included the FSF-published book Free as in Freedom, which is covered by the GNU Free Documentation License (GFDL). The FSF notes the GFDL permits reuse but argues that true compliance requires more than attribution: sharing complete training inputs, model weights, training configs and source code to preserve user freedom. The FSF warns it could seek such “freedom” remedies if it joins copyright litigation against Anthropic (e.g., Bartz v. Anthropic), emphasizing that while it rarely sues, it would demand sharing as compensation for GFDL violations. This raises broader questions about how permissive licenses apply to AI training and settlement terms for copyright claims.