What Questions Help in a Human or AI Game?

The follow-up matters more than the opener

A standalone question invites a standalone performance. A ladder makes the next question depend on the previous answer. That creates a small thread you can inspect for specificity, revision, and continuity.

Turing Test Live uses an active interrogator who questions two hidden witnesses before identities are revealed. The Human or Not research project likewise studied a live social game in which people chatted before deciding whether the other side was human or AI. The useful structural idea is simple: interaction creates evidence that a frozen paragraph cannot provide.

Do not hunt for one magic question. Build a short chain in which each answer changes what you ask next.

A four-step diagram showing ask, answer, follow up and reveal — Figure 1 · A useful question produces the next question; the final reveal remains the only known label.

Six useful question families

Specific memory. Ask for a small sequence, then request the detail that changed the outcome.
Constraints. Force a choice between two imperfect options, then ask which limitation mattered most.
Correction. Invite the witness to revise an earlier answer and explain what the first version missed.
Comparison. Ask for a concrete difference between two familiar things, then test the boundary.
Local context. Ask about something visible or established inside the current game rather than general knowledge.
Continuity. Introduce one detail, change topics, and return to that detail later.

Questions that add little information

Trivia, arithmetic, riddles, typos, response speed, and direct identity questions can all be imitated, looked up, delayed, or deliberately performed. Generic questions also invite generic answers. Use them for fun, not as proof.

A community discussion about Turing-style games repeatedly suggested returning to earlier details and asking follow-ups. That is useful as a source of player behavior, not scientific validation. The Question Lab turns that player instinct into a transparent three-step structure.

A comparison between a generic identity question and a three-step dependent question — Figure 2 · Strong questions create a thread to inspect; weak questions end where they started.

How to use the ladder fairly

Ask the same kind of question of every witness.
Keep the time and message limits comparable.
Judge the whole exchange instead of one spelling habit.
Write down what changed your mind before the reveal.
Keep the conclusion inside the game.

Six tiles for memory, constraints, correction, comparison, local context and continuity — Figure 3 · Rotate question families instead of repeating one favorite trick.

Sources and provenance

Turing Test Live (reviewed October 13, 2026: an interrogator questions two hidden witnesses before the reveal)
Jannai et al.: Human or Not? A Gamified Approach to the Turing Test (2023)
Reddit discussion: conducting a Turing-style test (community evidence for consistency follow-ups, not a controlled study)

All 18 ladders in the tool above were written for this page on October 13, 2026. They are prompts for a game, not excerpts from the sources.

This page is not a detector

No question proves that a witness is human or AI. The tool does not contact a model, analyze a reply, upload text, or assign an identity.