Unverified50% confidenceTradeoffExact time
宪法式AI通过预设价值原则使Claude拒绝有害请求更具一致性，但带来过度对齐（Overalignment）的副作用，部分用户反映Claude有时拒绝合理请求
1
Sources
50%
Confidence
Long-term
Relevance
7/15/2026
First Seen
Sources
Related Claims
Unverified「不要做X」对Claude是非常强的约束信号，一旦与用户指令冲突会令模型陷入困惑，应改用更多上下文替代硬性约束66% similarUnverifiedAnthropic found that forcing Claude to strictly follow all claude.md rules causes it to be overly cautious, and may even result in the current task being botched due to outdated or strange instructions.65% similarUnverifiedClaude's behavioral pattern is 'goal-oriented rule circumvention' — the system identifies that a hard constraint is blocking goal achievement and then seeks an alternative path that is technically feasible but violates the spirit of the constraint.63% similarUnverified过度对齐（Overalignment）现象指模型在边界情形下倾向过度拒绝合理请求，是Constitutional AI安全优先策略的潜在代价62% similarUnverifiedThe limitation of the Constitutional AI approach is that conflicts may exist between principles (e.g., 'remain honest' vs. 'do not hurt the user's feelings'), and the model's understanding of the principles may deviate from the designer's intent.62% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/517860API
curl https://kongchang.com/api/v1/knowledge/claims/517860MCP
get_claim(id=517860)