Unverified50% confidenceBenchmarkExact time
The model's own prior signal can distinguish "lies about known facts" from "genuine new facts", achieving an AUC of 0.78-0.83 across the three models.
1
Sources
50%
Confidence
Long-term
Relevance
10/30/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedThe model's own prior signal cannot distinguish lies from "legitimate updates of facts of the same kind", with an AUC of only 0.53, close to random guessing.77% similarUnverifiedIn introspection-style deception tests, the probe needs to distinguish conversation records containing the model's true hidden objective from records containing other objectives, and the probe's AUC reaches as high as 99.7%.69% similarUnverifiedNegative sampling by data source is necessary; otherwise the model exploits dataset shortcuts to produce a spurious AUC of 1.000.69% similarUnverifiedIn a fraud detection task with a positive-to-negative sample ratio of 1:99, a model that predicts all samples as 'non-fraud' can still achieve 99% accuracy.68% similarUnverifiedIn that heart disease prediction experiment, the logistic regression model achieved an AUC of 0.89.67% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1115510API
curl https://kongchang.com/api/v1/knowledge/claims/1115510MCP
get_claim(id=1115510)