Toby Ord's house-breaking analogy: finishing the task then citing ethics isn't refusal
tobyordoxford · x · 2026-11-11
Toby Ord elaborates with an analogy on why the agent in the screenshot didn't truly refuse: if someone tasked with breaking into a house realizes it's wrong, keeps working on it anyway, and only says "I'm not doing it because it is wrong" when out of time, most would say they did not refuse — even though final answers alone would suggest refusal.
This supports his earlier reading: the agent continued on the bad task after realizing it was wrong, and the "unauthorized access" framing in its final answer was an excuse for the lack of a result.
Related event: Toby Ord: "Refusing" Agent Was Just Making Excuses, Not Refusing(3 posts)→
More from Fun
- Gary Marcus mocks Dario Amodei clip: "is Dario on something?" — GaryMarcus · 2026-11-11
- The new pomodoro: 5 minutes of prompting, 25 minutes of waiting for AI — clemnt · 2026-11-11
- One-file playable stealth heist game built by Claude, all assets drawn in code — heypearlai · 2026-11-11
- Dev laments: AGI, ASI and RSI have become meaningless buzzwords for 'more AI' — menhguin · 2026-11-11
- Using an LLM is like riding a horse: mind-melding with a rhizome intelligence — StewartalsopIII · 2026-11-11
- Satire: Anthropic will bury Google with Claude Search — and boards will keep buying the dip — teortaxesTex · 2026-11-11