Bigger Than a Breadbox: A 20 Questions Game

How to start

1. Open it in your AI. Tap Open in ChatGPT or Open in Claude and press send, or paste the play prompt into any assistant that can open links. The chat must be able to open a link, because the Answerer reads a sealed file.
2. The cards load. The Answerer fetches the sealed file, reads you the start of the SHA-256 commitment, and opens the first thing with its class: animal, vegetable or mineral.
3. Ask, then name it. Ask yes/no questions and name the thing with "Is it a …?". Fewer questions score more; twenty is the limit.
4. Check the cards. After the tenth thing, check every logged answer against the sealed file, and the seal at aigamecards.com/verify.

Questions

Can ChatGPT play 20 questions?

Not reliably on its own. In a plain chat the thing you are guessing is not fixed anywhere you can check, so the AI can answer each question by what sounds likely and settle on a thing only when you guess. A 2026 paper proves that an agent restricted to the public conversation cannot both keep a secret unresolved and stay consistent with it. This card fixes the ten things first: each one and its answers to sixty-seven common questions are sealed under SHA-256 before your first question. The card tells the AI to fetch that file at the start, rule other questions by a written procedure and log each costed answer, so you can check them against the cards afterwards.

What does the seal prove, and what doesn't it?

It proves that the ten things and their coded answers were fixed before you asked anything; you can rebuild the hash at aigamecards.com/verify. It does not prove that every answer was right. Questions the codes do not cover are ruled by a written procedure from each thing's sealed description, and the AI can misapply a rule, so every answer is logged with the code or words it used, and the sealed file is yours to read afterwards. A score or log carried over from a save is the player's own record; the seal does not cover it.

Is this 20 questions game good for kids and car trips?

Yes. Every thing is one a child and a grandparent both know, nothing is gory, and nothing asks for a name or a personal detail. One person types and everyone asks, so it works as a car game or at the table. Ten things take an estimated forty-five minutes, about four or five minutes each, and a save code lets you stop for a break and carry on later. With younger children a parent should be at the keyboard.

What if my question isn't a plain yes or no?

The AI can also answer SOMETIMES (true of some ordinary ones, like a hat's colour) or PARTLY (true of a part, like a hammer's wooden handle); both cost a question. Questions that are not yes or no, name two or more things, add more to a name, or ask about the name come back REPHRASE, and trivia no card covers, such as who invented it or a date, comes back NOT ON ANY CARD; both are free. Examples after a description are ignored: "Is it a pet, like a cat or a dog?" is ruled as "Is it a pet?". The sixty-seven coded questions were chosen from the questions players asked in the Allen Institute for AI's twentyquestions dataset (48,638 questions from crowd-sourced games). Other questions are ruled by the written procedure, and each costed answer is logged.

Does the chat have to be able to open a link?

Yes. The card tells the AI to fetch the sealed file from aigamecards.com at the start and, if it cannot, to say so and stop rather than play; pasting the file yourself would show you all ten answers. A chat that ignores the card can still make things up, and the card marks and the sealed file let you catch that afterwards. aigamecards.com/launch-matrix lists which apps can currently open links.