Anthropic's fable 5.5 model faces scrutiny over hidden safety checks

Anthropic's Fable 5.5 model, a next-generation large language model, has drawn attention for its advanced intent understanding and safety mechanisms. Users report the model excels at interpreting complex tasks and user intent, addressing underlying issues beyond surface-level instructions. A key feature is the 'Self-Assessment' mechanism, which runs a hidden safety check after each response. This system, reusing existing context and KV Cache, identifies 78% of unsafe interactions and detects high-risk behavior at 93%. The process uses less than 2% of total inference costs, with no impact on user billing or latency. Anthropic claims the system respects privacy, though users express concerns over implicit 'moral judgment' and potential surveillance. The Self-Assessment system operates without external models, relying on the same AI that generated the response. It processes a few dozen tokens, outputting a 10-token JSON result with no data retained for individual users. Anthropic uses aggregated risk signals for offline analysis, aiming to uncover new threats like prompt injections. Despite its efficiency, the mechanism has sparked debate over its impact on user trust and freedom of expression. The company emphasizes safety audits are platform-funded and do not penalize users, though critics argue the system creates an invisible sense of oversight.