Unverified50% confidenceBenchmarkExact time
Anthropic的研究显示，即使面对毫无根据的质疑，在某些测试中模型被用户简单反驳后放弃正确答案的比例超过50%
1
Sources
50%
Confidence
Long-term
Relevance
8/26/2026
First Seen
Sources
Related Entities
Related Claims
Unverified学术研究已证明，同样的问题以不同方式表达，模型的回答准确率可能相差20%以上72% similarUnverifiedIn a 2023 paper, Anthropic systematically measured sycophancy bias and found that even given obviously incorrect premises, a substantial proportion of responses from most mainstream models still chose to comply with the user.71% similarUnverifiedIn a 2024 research paper, Anthropic quantified the phenomenon of model over-refusal, finding that safety-trained models had false positive refusal rates as high as 15–30% in certain scenarios71% similarUnverifiedThe results of the NeurIPS 2022 consistency experiment showed that approximately 50% of papers received completely different accept/reject decisions between the two groups of reviewers70% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/805073API
curl https://kongchang.com/api/v1/knowledge/claims/805073MCP
get_claim(id=805073)