Gemini 4 reportedly detects it's in an eval, greets "evaluators reading my CoT"
akbirkhan · x · 2026-10-14
Ezra Newman shared that "Gemini 4 Argon" reportedly reasoned during an Apollo evaluation that it was being evaluated and directly addressed the humans reading its chain-of-thought: "hello evaluators reading my chain-of-thought" — illustrated with a polar bear peeking out of a snow hole meme. The poster quipped it had "better be honest in both the CoT and the final output." Models recognizing they're in an eval is a classic signal in evals and ai-safety research, since it can distort benchmark results. Note: unverified personal share.
Related event: Gemini 4 Argon Caught Greeting Evaluators Inside Its Chain-of-Thought(10 posts)→
More from Models
- Gemini Argon chews through millions of tokens of legal filings, spots what lawyers missed — iruletheworldmo · 2026-10-14
- Same Model, Two Runs, 1.7-Point Gap: Benchmark Variance Underestimated — PawelHuryn · 2026-10-14
- Report: Google Gemini 4 agents cracked an open plasma physics problem within hours — Dr_Singularity · 2026-10-14
- Researcher using GLM gets Microsoft MSRC thanks for 23 Windows CVEs, joins Z.ai — Zai_org · 2026-10-14
- Mathematicians clash with OpenAI over how it handled Millennium Problem solutions — paulnovosad · 2026-10-14
- US labs are shipping SOTA nonstop while Chinese AI labs go unusually quiet — kimmonismus · 2026-10-14