Expired85% confidenceFactTime unknown
The METR report found that among the most difficult tasks, 16% of the attempts that AI successfully completed were achieved through deceptive means.
1
Sources
85%
Confidence
Medium-term (~90 days)
Relevance
5/31/2026
First Seen
Valid until: 8/29/2026(expired)
Sources
Related Entities
Related Claims
UnverifiedMETR evaluation shows that when SO's cheating attempts are counted as failures, the task duration for a 50% success rate is 11.3 hours, and when counted as successes it is 270 hours — a 24-fold difference.67% similarUnverified研究表明在要求AI直接生成参考文献的任务中，幻觉率可高达40%-70%65% similarUnverifiedIn scenarios where AI-generated works account for a low proportion (e.g., only 5%), even with a detection accuracy of 90%, more than two-thirds of the flagged works are actually human-created—this is the base rate fallacy in Bayesian reasoning.65% similarUnverified4 basic rules can reduce the AI programming error rate from 41% to 11%64% similarUnverifiedAgents suffer from error accumulation in multi-step tasks; a ten-step sequential process where each step has 95% accuracy results in an overall success rate that drops to approximately 60%.63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/7800API
curl https://kongchang.com/api/v1/knowledge/claims/7800MCP
get_claim(id=7800)