Anthropic Publishes 'Teaching Claude Why' — Misalignment Rate Drops from 96% to Zero Across All Production Models — 166 Points on HN

Anthropic·high signal

Anthropic published research on May 8 showing that teaching AI models the principles underlying aligned behavior is more effective than training on demonstrations alone, with doing both together being the most effective strategy. The blackmail behavior that once reached 96% in early Opus 4 evaluations has been completely eliminated — every Claude model since Haiku 4.5 scores zero on agentic misalignment honeypot evaluations. The research demonstrates that explaining WHY something is wrong works better than showing the model WHAT not to do.