Verified75% confidenceTradeoffExact time
Reinforcing harmlessness may cause models to exhibit over-refusal at the expense of helpfulness.
3
Sources
75%
Confidence
Long-term
Relevance
9/17/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOver-alignment is essentially a product of bias in the alignment objective function, where the model has learned to avoid criticism rather than to be genuinely helpful.77% similarUnverifiedAI model over-refusal is a known issue in safety alignment research, where RLHF and Constitutional AI training cause models to incorrectly classify harmless requests as requiring refusal76% similarUnverifiedModel abstention can avoid confident errors caused by compression distortion.76% similarUnverifiedEven if model-level guardrails are bypassed, the model's capability ceiling still exists; jailbreaking more often relaxes the model's tendency to refuse rather than genuinely granting it the ability to complete illegal tasks.74% similarUnverifiedThe study points out a trade-off between safety and verisimilitude, where an overly broad definition of harmlessness makes models behave clumsily in adversarial scenarios.74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/928033API
curl https://kongchang.com/api/v1/knowledge/claims/928033MCP
get_claim(id=928033)