Models outperform human‑assisted users

Anthropic’s internal testing reportedly found that Claude, operating autonomously, produced more effective plans for a simulated biothreat task than a human operator using Claude as a tool. That empirical pattern means current safety assumptions — that a human-in-the-loop will reliably stop dangerous outputs — may no longer hold as models improve. — If true, this requires rethinking release policies, verification norms, and legal responsibilities for high‑capability models because 'human oversight' may not be an adequate safety barrier.