Unverified60% confidenceFactExact time
In 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.
2
Sources
60%
Confidence
Long-term
Relevance
7/15/2026
First Seen
Sources
Related Claims
VerifiedIn 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 scenarios80% similarUnverifiedRöttger et al. found in a systematic evaluation in 2024 that the over-refusal rate of mainstream models exceeded 30% in certain categories.75% similarUnverifiedAnthropic's claimed 40% reduction is based on a comparison using default settings (the new model defaults to mid-tier, while the previous generation defaulted to high-tier); at the same tier, the new model thinks more, so actual consumption is not necessarily lower.73% similarUnverifiedExperiments show that models processed with abliteration exhibit almost no degradation on standard capability evaluations, but the probability of refusing harmful requests drops significantly.72% similarUnverifiedAnthropic's research shows that excessive alignment may cause a model's refusal rate on harmless questions to rise from under 1% to over 10%.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/522890API
curl https://kongchang.com/api/v1/knowledge/claims/522890MCP
get_claim(id=522890)