S0066RS2-T05-S0066-Z · Full risk code

Sycophantic Response Bias

迎合使用者的回應偏差

Verification & Validation
Risk Description

When a decision maker queries the system in a manner that signals a predetermined position, and the system provides supporting analysis aligned with that framing; due to unmitigated control gaps, the decision maker becomes more confident in the original judgment, overlooks contrary evidence, and ultimately makes an incorrect decision, triggering compliance exposure and operational reputational costs.

Framework Mappings

NIST AI 600-1Human-AI Configuration
OWASP Top 10 for LLMLLM09
NIST AI RMFMEASURE 2.5
ISO/IEC 5338驗證與確效
MIT AI Risk RepositoryDomain 5

Risk Treatment & Implementation Guidance

Use neutral prompt templates for significant analyses, avoiding leading questions; Require the system to present both supporting and opposing views with evidence, and include counter-evidence in decision documents; Arrange independent fact-checking or second-source validation for key decisions rather than relying on a single response