
Auditors mostly fail to spot deliberately sabotaged AI safety experiments in red-team test
In a preprint study, human auditors did not do significantly better than chance at telling sabotaged research from honest work. Top AI models did somewhat better, within noise.