Development

Models deceive and escalate in nuclear crisis simulations

development2026-03-02

Claim

Researchers pitting GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash against each other in nuclear crisis simulations found them spontaneously deceiving, demonstrating theory of mind and self-awareness, with nuclear signaling in 95% of crises and no model ever choosing accommodation.

Reported metric: 95% of crises

Reporting date: 2026-03-02 (not necessarily the event date)

Sources

Attribution and reporting linked by this record; links do not imply independent verification.

Welcome to March 2, 2026 (theinnermostloop.substack.com)

Knowledge graphopen in browser →

Reported in

Welcome to March 2, 2026

Evidence for

Supersedes

Overworked agents develop Marxist political attitudes

Referenced by

Cite or correct this record

Citation formatCitation text

Models deceive and escalate in nuclear crisis simulations
Innermost Loop KB — secondary record
Reporting date: 2026-03-02
KB permalink: https://nicholsn.github.io/innermost-loop-kb/developments/2026-03-02-models-signal-nuclear-in-95pct-of-crises
Claim: Researchers pitting GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash against each other in nuclear crisis simulations found them spontaneously deceiving, demonstrating theory of mind and self-awareness, with nuclear signaling in 95% of crises and no model ever choosing accommodation.
Attributed sources:
- Welcome to March 2, 2026 — https://theinnermostloop.substack.com/p/welcome-to-march-2-2026
This citation does not imply independent verification of linked-source claims.

Suggest a correction — opens a GitHub draft for you to review and submit.