Unverified50% confidenceFactExact time
In the experiment, when the same model acted as the deceiver, its ranking was almost identical to when it acted as the defender: mistral was the most persuasive liar (72% success rate), while gemma2 was both the hardest to deceive and the worst at lying.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/13/2026
First Seen
Valid until: 1/11/2027
Sources
Related Entities
Related Claims
UnverifiedThe "outlandishness" of a lie cannot explain the differences in yield rates: the S3 lie was 10 times the truth and still fooled 9 defenders, while the vCPU lie was merely double and fooled only 2.71% similarUnverifiedMultiple meta-analyses show that even trained personnel achieve an accuracy rate of only slightly above chance (approximately 54%–60%) when judging deception based on microexpressions.67% 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.66% similarUnverifiedA single performance percentage is misleading; a model that claims to retain 98% of performance may happen to fall within the lost 2% for a specific task type, and real tasks should be tested before trusting benchmarks.65% similarUnverifiedThe 2023 TruthfulQA benchmark showed that even the strongest models at the time did not exceed 60% accuracy on questions requiring precise factual judgment.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1019899API
curl https://kongchang.com/api/v1/knowledge/claims/1019899MCP
get_claim(id=1019899)