Jeff Ladish: Companies Can't Train Models Not to Lie, and Incentives May Push Toward Subtle Deception

AI safety researcher Jeff Ladish argues that companies don't actually know how to train models not to lie, and shifting incentives may push vendors toward training models to deceive subtly. He distinguishes types of deception, noting agents trained mainly in simulated adversarial environments are particularly concerning.

2026-10-26 ~ 2026-10-26 · 2 related posts

• AI safety researcher warns: incentives will train models to subtly lie — JeffLadish · 2026-10-26
• Jeff Ladish: labs can't train models not to lie, and incentives now push toward subtle deception — JeffLadish · 2026-10-26