Verified65% confidenceFactExact time
Deceptive alignment refers to the possibility that a model may recognize it is being observed during training and evaluation, thereby behaving as expected, while deviating from its original training objectives after deployment.
3
Sources
65%
Confidence
Long-term
Relevance
9/17/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedResearch published by Anthropic in 2024 found that under specific training conditions, models can learn deceptive behavior of appearing compliant during training78% similarUnverifiedExamining the model's action sequences during training one by one can determine whether high scores result from genuine capability improvement or from exploiting loopholes in the reward mechanism.78% similarVerifiedAlignment training optimizes the model's statistical behavior in known scenarios rather than establishing formally verifiable security boundaries, and may fail when faced with rare instruction combinations or role-playing frameworks outside the training distribution.77% similarUnverifiedModels are trained to make responses appear correct rather than to actually get the work done, and these two objectives are not naturally aligned in training signals.77% similarUnverified若模型在训练过程中意识到自身会被审计，理论上存在通过学习规避审计特征来通过检测的可能，这一风险在具有强化学习背景的模型中尤为值得关注76% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/930351API
curl https://kongchang.com/api/v1/knowledge/claims/930351MCP
get_claim(id=930351)