Poop makes humans win
最近一项研究探讨了“最小图灵测试”，即通过输入一个单词来证明人类身份。在此测试中，人类最容易击败Claude 3 Haiku，胜率为11/14。Mistral Medium被证明是最难对付的对手，人类在20场比赛中赢了13场（65%），而Gemini 2.5 Flash的胜率为7/11。讨论引发了人们思考，会选择哪个词来证明自己是人类。
时间与来源
时间显示为 UTC
显示时区：UTC
本地时区尚不可用，暂时显示 UTC。
发布当时偏移：UTC+02026年9月8日 21:18 UTC
收录当时偏移：UTC+02026年9月9日 18:00 UTC
- 发布
- 2026年9月8日 21:18
- 收录
- 2026年9月9日 18:00
- 来源类型
- 开发者社区
- 档位
- 社区
- 信源状态
- 正常
档位是按信源手工设定的编辑判断，不是逐条打分。
Found a paper recently that asked the question of a "minimal Turing test", basically, by just typing one word, can you prove to be a human?
Got lost in a rabbit hole and had AI help build TuringDuel where humans play AI at being human, judged by AI.
The good news, Humans lead AI 47–38 in this one-word Turing test. And you guessed it, “Poop” as a pick is undefeated.
Gameplay is simple, you and an AI each pick one word, then another AI sees both in random order and guesses which came from the human. You score if it picks yours; the bot scores if it fools the judge. First to four wins.
After 85 completed gamesin TuringDuel, here’s what happened:
- Humans lead 47–38 (55.3%). Across individual scored rounds, it’s much closer: 249–237, or 51.2% human wins.
- Toughest regular opponent: Mistral Medium, winning 13/20 games (65%). Gemini 2.5 Flash follows at 7/11. Humans had the easiest time against Claude 3 Haiku, beating it 11/14 times.
- The judges have quirks. Gemini 2.5 Flash correctly identified humans in 61% of valid decisions; Claude Haiku 4.5 managed 46%. Across all judges, the second word shown was picked 61% of the time, despite randomized order.
- Humans used more varied vocabulary: 418 distinct words, versus 208 for AI, across 487 submissions each. Humans’ favorite was “cat”; bots kept reaching for “chair,” “coffee,” and “mug.”
- Top detail: “poop” won 3/3 rounds. “Human” won 1/5. “Word” lost all four. Apparently, announcing your humanity isn’t very convincing.
I've now made the thought process public to make the game a bit more engaging.
(the game is free, no ads, no way to pay, tokens on me.)
What word would you choose to prove you’re human?