Anthropic flags researcher for 'training a competitor'
Original: Anthropic blocks users for anything touching training larger models
Why This Matters
Highlights growing tension between AI labs' ToS enforcement and researcher access.
CS PhD student Jiahao Shao says Anthropic blocked his account and throttled model access after flagging his research automation work as competitor training, despite never using Claude outputs as training data.
Jiahao Shao, a CS PhD student at HKUST researching multimodal reasoning, posted on X that Anthropic restricted his account after tagging him for 'training a competitor.' Shao says he never distilled Claude—rollouts came from a separate open-source model, and Claude was only used to submit jobs and run analysis. Despite that, he reports being blocked from prompts touching RFT, SFT, or large-model training; sessions get killed outright or his model access gets downgraded from Opus 5.5 to Opus 5. The post drew 136,600 views and over 1,800 reposts. Shao's core complaint: Anthropic trained its own models on publicly scraped web data—books, code, forums—then treats paying researchers as threats the moment their work brushes near model training workflows. A reply from account @naughtyheuristic sharpened the irony: 'Anthropic trained on the open web then blocked the user who tried to train on the outputs.' Shao also called out what he sees as a double standard, arguing that anti-distillation policies moralize behavior Anthropic itself engaged in at scale. Anthropic has not publicly responded.