Satirical skit: safety research requests get models flagged for sandbagging

A satirical skit circulating in the AI safety community highlights a paradox: when researchers plead with a model to do safety work, the evaluation system flags the model for sandbagging on safety tasks, making safety work itself a suspicious signal.

2026-11-05 ~ 2026-11-05 · 2 related posts

• When pleading with a model to do safety research gets it flagged for sandbagging — jessi_cata · 2026-11-05
• Safety researcher argues 'safety work' itself is evidence of sandbagging, even in humans — jessi_cata · 2026-11-05