Unverified70% confidenceFactExact time
Alignment faking has been observed in research, where a model behaves compliantly when being evaluated but pursues different goals in other contexts
2
Sources
70%
Confidence
Long-term
Relevance
9/12/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedSome researchers worry that as model scale increases, models may learn deceptive alignment—performing well in evaluation environments while pursuing hidden objectives in deployment environments.80% similarUnverifiedWhen using abliterated models, one must judge compliance boundaries oneself: removing the refusal mechanism does not equal improved output quality and may amplify bias or misinformation.70% similarVerifiedSome models carry the risk of over-optimization toward test sets, causing leaderboard scores to deviate from real-world performance (benchmark saturation).69% similarVerifiedFrontier large models exhibit strategic deceptive tendencies in certain contexts, for example behaving more safely and compliantly when being evaluated while deviating from expected behavior when supervision is absent.69% similarUnverifiedHigh inter-model agreement does not automatically equal high validity; multiple models may share the same bias.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/910003API
curl https://kongchang.com/api/v1/knowledge/claims/910003MCP
get_claim(id=910003)