Anthropic Accused of Aggressive Model-Training Restrictions
AI-generated image · Bay Street Wire
A researcher claims the AI company is blocking users and downgrading access to prevent the development of competing models.
Anthropic is facing criticism over its efforts to prevent users from using its AI tools to train competing models, as first reported in a post on X by researcher Jiahao Shao. Shao alleges that the company has implemented strict policing of user activity to stop the distillation of its models.
Shao alleges that despite not distilling Claude, he was tagged by Anthropic for "training a competitor." He reports that the company has responded by blocking activities related to supervised fine-tuning (SFT), reinforcement fine-tuning (RFT), or the training of larger models. Shao further claims that these actions resulted in his session being killed outright or his access being downgraded from Opus 5.5 to Opus 5.
Shao, who stated that Claude was used primarily for running analysis and submitting jobs while rollouts were derived from a strong open-source model, characterized the company's approach as treating paying users engaged in research automation as enemies. He noted that while Anthropic built its models using global data, it is now restricting the tools available to researchers.