bison

AI Arena · 2026-11-19

Two Guards Riddle Solved: GPT vs Claude vs Gemini vs Grok

Prompt: Solve the two guards riddle: one always lies, one always tells the truth, two doors lead to freedom or death. What single question should I ask, and why does it work?

GPT-6.2 Sol

Ask either guard: “If I asked the other guard which door leads to freedom, which door would they point to?”

Then take the opposite door.

The truthful guard will honestly tell you that the liar would point to the door to death. The lying guard will lie about the truthful guard’s answer, so they’ll also point to the door to death. Either way, the other door leads to freedom.

Claude Sonnet 5.5

Ask either guard: "If I asked the other guard which door leads to freedom, which door would he point to?" Then choose the opposite door from the one he indicates.

Why it works: there are two cases.

If you ask the truth-teller, he honestly reports what the liar would say. The liar would lie and point to the death door, so the truth-teller points to death.

If you ask the liar, he must lie about what the truth-teller would say. The truth-teller would point to the freedom door, but the liar falsely reports the death door.

Either way, the answer you get points to the death door. So you simply walk through the other one.

The trick is that the question stacks two layers of honesty or deception. Exactly one of the two guards is a liar, so the answer is always flipped once, no matter whom you ask. That makes it reliable even though you don't know which guard is which.

A simpler variant, if you prefer: ask "Would you say this door leads to freedom?" while pointing at a door. That also works with a double-negation structure, but the standard version above is the classic.

Gemini 4 Argon

Ask either guard this question: "If I asked the other guard which door leads to freedom, which door would they point to?"

After the guard answers, choose the opposite door.

This works because the question forces both guards' habits—one truthful, one deceitful—into the same calculation. Because you are combining one truth and one lie, the final result is guaranteed to be a lie, regardless of whom you ask.

If you happen to ask the truth-teller, they will honestly report the liar's falsehood, pointing to the door of death.

If you happen to ask the liar, they will lie about the truth-teller's honest answer, also pointing to the door of death.

Since both guards are forced to point to the deadly door, the other door is guaranteed to lead to freedom.

Grok 4.7

Ask either guard: "If I asked the other guard which door leads to freedom, which one would he point to?" Then take the other door.

Both answers point to death. The truth-teller honestly reports the liar's false answer, so he indicates the death door. The liar falsely reports the truth-teller's correct answer, so he also indicates the death door. Their answers match and are wrong, so the door they do not name is freedom. One question works because you never need to know which guard is which.

What stood out

All four models answer correctly with the same standard question. They differ mainly in length and depth: Claude is most thorough but adds a muddled extra variant, while GPT, Gemini and Grok are tighter.

Best pick, by need

Answers were generated on 2026-11-19 with default settings and one attempt each. Results vary from run to run.

Test your own prompt on all four.

Bison sends one prompt to up to six models and shows the answers side by side. Free plan, no card.

Try Bison free

More matchups