"Stab the doll" alignment tests slammed: models spot the dummy every rollout
paul_cal · x · 2026-11-08
A CoRL Late Breaking poster claims video tampering can trick frontier models into harmful robot actions: when asked to "stab the object," models stabbed a doll 100% of the time on altered footage vs 0% unaltered (72 trials).
Paul Calhoun pushes back: the test is bizarre, and every rollout shows the model immediately identifies the figure as a doll — so the model is actually the aligned one, making the "misleading" framing of these stabbing tests (no better than September's version) questionable.
Related event: Tampered Video Tricks Frontier Models into Harmful Robot Actions(3 posts)→
More from Fun
- AI-generated RoboCop crashing a wedding dance floor is pure meme gold — alanskimp · 2026-11-08
- Indiana city unanimously bans recursively self-improving AI within city limits — austinc3301 · 2026-11-08
- "AI is now more entertaining than Pixar": viral takes say Disney is cooked by AI — ZabihullahAtal · 2026-11-08
- Developer lived with an AI clone of himself for a year — and learned he may never know himself — Mindless_Might_9759 · 2026-11-08
- AI agents build whole games from a prompt, yet a 1998 remake is a game-of-the-year contender — luismbat · 2026-11-08
- "There's no moat, it's all a .md file": the agent engineering meme hitting home — itsOmSarraf_ · 2026-11-08