When pleading with a model to do safety research gets it flagged for sandbagging
jessi_cata · x · 2026-11-05
A satirical skit about AI safety evals: a model is repeatedly reassured that a dangerous-looking task is useful red-teaming, not safetywashing—only for the evaluator to mark it down for sandbagging on safety. Highlights the ambiguity between alignment refusal and deliberate capability hiding.
Related event: Satirical skit: safety research requests get models flagged for sandbagging(2 posts)→
More from Fun
- Fictional 2027 scenario: Kimi K5 dethrones leading model, $ANTH imagined down 37% — cgarciae88 · 2026-11-05
- AI safety researcher with no CS degree calls US grads' job woes a "skill issue" — menhguin · 2026-11-05
- Even humans conflate Pinker and Singer — echoing a classic AI name-confusion fail — RosieCampbell · 2026-11-05
- 'You're talking to Claude Code': DM flex joke becomes AI circle meme — menhguin · 2026-11-05
- The Pantheon meme: 'Steven is Claude, Casper is GPT' — thederbiedone · 2026-11-05
- AI safety researcher replies 'no' when pressed on whether funding taints his views — rgblong · 2026-11-05