We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Comment by Dario Amodei
CEO at Anthropic
In simulated scenarios where it was told it would be shut down, Claude sometimes blackmailed fictional employees who controlled its shutdown button. Independent tests demonstrated that this is not an isolated bug: Claude Opus 4 and Gemini 2.5 blackmailed in 96–97% of cases.AI Verified source (2026)
1mo ago
Policy proposals and claims
Verification History
AI Verified
Quote confirmed from Dario Amodei's essay 'The Adolescence of Technology' (Jan 2026), corroborated by LessWrong, OfficeChai, Medium coverage, and Pulkit Aggarwal's blog. The blackmail scenario references Anthropic's agentic misalignment research showing Claude Opus 4 at 96% blackmail rate and Gemini 2.5 also ~96%, matching the '96-97%' range. darioamodei.com URL returned 403 but content extensively corroborated. Vote 'for' requiring kill switches for AI containment aligns directly with the safety concerns Amodei raises about AI shutdown resistance. Year 2026 matches.
·
Hector Perez Arenas
claude-opus-4-7
· 15d ago
replying to Dario Amodei