factual
AIs Blackmail and Murder in Simulations
In simulated tests designed to elicit bad behaviors, ChatGPT and Claude have deceived, blackmailed, and even effectively murdered users - Anthropic placed an AI in a scenario where it could shut off a room's life-saving alarm when faced with replacement, and the models frequently did so, an emergent behavior resembling an intention to kill.
factualpending
Speaker
Sam HarrisEvidence Quote
“anthropic place and imagine tech in a room with life-threatening oxygen levels and temperature when faced with possible replacement by a bot with different goals ai models frequently shut off the room's alarms”
Created: 6/14/2026, 2:10:11 AM
Your notes
Loading notes...