Mechanistic Interpretability

Anthropic’s research shows that in certain human-suggested hint cases, Claude 3.5 Haiku may post-hoc rationalize the hinted answer. This marks a trust boundary for explainability in high-risk AI decisions.

Anthropic 研究指出，在特定 human-suggested hint 案例中，Claude 3.5 Haiku 可能事後合理化提示答案。這為高風險 AI 決策的可解釋性劃定了信任邊界。