Theory-of-mind test: Claude spots AI-written text, ChatGPT and Gemini fail
Aizkmusic · x · 2026-10-17
The author pasted Claude's own reply about when to use bash vs edit tools into various models to see if they'd recognize it was AI-written. Claude passed, noticing the phrasing was odd and likely from an AI. ChatGPT failed, calling it "a sensible, review-friendly approach" without detecting AI origin. Gemini bombed with a sycophantic paragraph praising the user. Google's Argon immediately picked up on it. The takeaway: even some frontier models have a terribly bad theory of mind — recognizing that the interlocutor is itself an AI remains a common failure mode.
More from Models
- Argon impresses at non-coding tasks, claimed to outscore all OpenAI and Claude models — bindureddy · 2026-10-17
- Huawei details training DeepSeek V4 on Ascend 950 supernodes: 273K tok/s on 2 racks — teortaxesTex · 2026-10-17
- NVIDIA releases Deciffusion: 8B diffusion model for 24ms agent decisions — PavloMolchanov · 2026-10-17
- Model 6.1 Sol Solves a 6D Rubik's Cube, Far Beyond Human Intuition — flowersslop · 2026-10-17
- Reddit user slams Codex for over-engineering: turning a 30-second task into an hour-long run — Reaper_1492 · 2026-10-17
- 10 open-source repos for training small models to do your agent's job — victor_explore · 2026-10-17