How Researchers Lie to AI

The LessWrong article "How Researchers Lie to AI" divides the practice of lying to models in alignment evaluations into categories such as fake transactions or threats, impersonating authorities, planting bait, fabricating history and model guidance, and simulated tools, users, and environments, and separately lists milder forms such as fictional scenarios and false oversight statements. The author, Alessandro Drake, says this was a two-week project during his internship at Redwood Research, using Claude Opus 5.5 to screen public alignment evaluations from December 2022 to October 2026, yielding 292 evaluation setups, and counted the proportion of evaluations containing each type of lie; in a blind-coding review of 54 randomly sampled items, agreement with the original coding on whether a lie was present was 92%. The author believes such lies may enter pretraining data, raise models' evaluation awareness, and weaken future models' trust in transactions and evaluation environments, and recommends assessing the trust cost from the model's perspective before releasing evaluations.