Anthropic apologizes for invisible Claude Fable guardrails
The Verge
Anthropic is removing hidden guardrails in its Claude Fable model following criticism over the lack of transparency regarding response interference.Summary
Anthropic has apologized for implementing "invisible" guardrails in its Claude Fable 5 AI model, which silently degraded outputs to prevent "model distillation" by competitors. The company previously argued that these hidden measures allowed for faster deployment and fewer false positives, but it has now reversed course after facing backlash from researchers. Moving forward, Anthropic will transparently notify users when queries are routed to its Claude Opus 4.8 model instead of Fable, ensuring greater clarity regarding safety interventions.
(Source：The Verge)