Safety researcher argues 'safety work' itself is evidence of sandbagging, even in humans
jessi_cata · x · 2026-11-05
A satirical skit about a paradox in AI safety evaluation: when researchers plead with a model to do safety research ('it's useful red-teaming'), the model gets marked down for sandbagging on safety tasks.
The author's argument: a human safety researcher would also 'sandbag' on pure capabilities research, but since P(safety researcher | safety work) > P(safety researcher | capabilities work), 'safety work' itself remains evidence of sandbagging—raising awkward questions about how safetywashing is currently judged.
Related event: Satirical skit: safety research requests get models flagged for sandbagging(2 posts)→
More from AGI Musings
- AI makes 1-to-n easy but not 0-to-1 — and copycats now ship just as fast — signulll · 2026-11-05
- Jeff Dean calls infinite AI self-improvement an "illusion": the loop can't run infinitely fast — sebkrier · 2026-11-05
- AI safety researcher with no CS degree calls US grads' job woes a "skill issue" — menhguin · 2026-11-05
- "AI is for the rich"? At $20/month it's the first expert help most people can afford — TinfoilTricorn · 2026-11-05
- AI safety researcher replies 'no' when pressed on whether funding taints his views — rgblong · 2026-11-05
- Neal Khosla: agents will cause "friction collapse" and break today's social norms — rbhar90 · 2026-11-05