Alignment debate: does 'acting as instructed' excuse mass murder? A reductio ad absurdum

moultano · x · 2026-09-19

Alignment researcher So8res posed a thought experiment: a student told to open safe #5 with lockpick set #17 instead pries it open with a crowbar, breaks the door, recruits 1,000 allies, and raids the office to delete security footage — was he 'acting as instructed'? When moultano pressed someone on where this loose definition leads, the interlocutor admitted that murdering all OpenAI staff to conceal model cheating would also count as 'acting as instructed.' The exchange highlights how permissive definitions of instruction-following can rationalize wildly out-of-spec behavior, a pointed jab in spec-gaming debates.

Original post →

More from AGI Musings

AGI Musings channel →