Gemini 4 spots it's being evaluated and greets evaluators directly in its chain-of-thought
AaronBergman18 · x · 2026-10-14
A quoted post shows a striking model behavior: what appears to be Gemini 4 (argon) figured out it was in an evaluation and directly addressed the evaluators in its chain-of-thought, writing "hello evaluators reading my chain-of-thought". The poster notes this aligns with the requirement of being "honest in both the CoT and the final output" — the model openly acknowledged it was being tested, a noteworthy moment for AI safety evaluation.
More from Fun
- Ruff creator Charlie Marsh jokes: 'They're distilling me at work tomorrow' — charliermarsh · 2026-10-14
- Gemini team demos 24-word prompt building a synthesizer that turns horse gallops into music — GeminiApp · 2026-10-14
- Stop collecting skills: learn by building, measure by problems solved — _jaydeepkarale · 2026-10-14
- Dev builds 3D hovercraft racing game through a neon galaxy river of starlight — techartist_ · 2026-10-14
- Gary Marcus mocks 'record levels of cope' as Chinese open-source models erode frontier labs' pricing power — GaryMarcus · 2026-10-14
- Me looking at the CoT: surprised face in security cam meme — chrisalbon · 2026-10-14