One-word Turing test: humans lead AI 47-38, and 'poop' is undefeated
jacob-indie · reddit · 2026-09-09
Inspired by a 'minimal Turing test' paper, a developer built TuringDuel: a human and an AI each pick one word, another AI judges which came from the human, first to four round wins takes the match. Across 85 completed games:
- Humans lead 47-38 (55.3%); per-round it's closer at 249-237 (51.2%).
- Hardest opponent: Mistral Medium (65% win rate), then Gemini 2.5 Flash; easiest was Claude 3 Haiku.
- Judge quirks: Gemini 2.5 Flash identified humans 61% of the time, Claude Haiku 4.5 only 46%; all judges favored the second word shown 61% of the time despite randomization.
- Humans used 418 distinct words vs 208 for AI; top human word was 'cat', bots kept choosing 'chair', 'coffee', 'mug'.
- 'poop' won 3/3, 'human' went 1/5, 'word' lost all four — claiming humanity is the least convincing move.
More from Fun
- Samsung really sells a "Bespoke AI Kimchi Plus" fridge, complete with kimchi cartons — airkatakana · 2026-10-26
- User Claims Codex Autonomously Reverse-Engineered Serum2 and Stripped Its License — Kyrannio · 2026-10-26
- ChatGPT Memory Trick: One Prompt to Uncover Your Ideal Partner Profile — Kyrannio · 2026-10-26
- Instagram Is Flooded With 'Opus 5.5'-Made Cave Incident Videos — Angaisb_ · 2026-10-26
- Solo dev renders 52-shot cinematic game trailer with Claude Code on one laptop for a few dollars — freechoice · 2026-10-26
- Quack Shot Starring Donald Duck gets a native C++ port, shown compiling in Visual Studio — ssh4net · 2026-10-26