Under Caution: An Interrogation 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. No paid plan required.
2. The interview room opens. The game master loads the sealed file, reads you its commitment and prints the first suspect's signed statement, the exhibits and the topics.
3. Ask and cite. Ask about a topic, or cite two IDs, such as a statement line and an exhibit, that cannot both be true. Any message with two IDs is a citation.
4. Confession and verify. Win the admissions on the sheet and the suspect confesses. Afterwards, open the reveal and rebuild the seal at aigamecards.com/verify.

Questions

Can ChatGPT play an interrogation game?

It can play one, but on its own it plays the suspect loosely: with nothing written down, a chat suspect can crack too easily when you push, or invent a detail to escape. This card fixes every answer the suspect gives, and the list of contradictions that break each lie, before your first question, and seals them under SHA-256. By the card's rules only a cited, listed contradiction moves the suspect; threats, promises and bluffs win nothing. The seal proves the answers and the list did not move. It does not prove every ruling was right: the reveal prints both, so you can check each one afterwards.

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

It proves that every answer, admission, confession and listed contradiction was fixed before your first question: paste the reveal into aigamecards.com/verify and the commitment rebuilds in your browser. It does not prove the AI ruled every citation correctly. In play you can challenge your last citation; afterwards you can test every "So? That proves nothing." against the list and the sheets yourself. Before release, every pair of items in all three interviews was checked against a timeline of each case: every contradiction found is either on the list or uses a claim the suspect has already taken back, which gets a free "Already admitted." Saves are not signed, so a score resumed from a save is on your honour.

How long is it, and who is it for?

Three interviews of about ten minutes each, for one player. It is rated 12+: an insurance arson, a drink-driving hit-and-run that breaks a teenager's wrist, and a fatal fall during an argument, told plainly and without gore. "Under caution" is the British term for a police interview after the suspect has been told they need not say anything; each suspect here has been cautioned and has declined a solicitor (a lawyer). The card tells the game master to refuse threats and promises of leniency; by its rules they never earn an admission. The card asks for no personal details.

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

Yes. The card tells the game master to fetch the sealed interview file from aigamecards.com when the game loads, to rule from it, and not to start if the file will not open; pasting the file yourself would show you every answer. aigamecards.com/launch-matrix lists which apps can currently open links.

Can ChatGPT run a detective game?

Yes, if the solution exists before you start. On its own, a chat model keeps no private notes between replies, and a 2026 paper proves that such a model cannot guarantee to keep an answer both hidden and fixed (arXiv 2601.06973, tested on hangman and on a model playing a patient with a hidden diagnosis). This card writes every interview's truth and every suspect line down first; the card tells the AI to answer only from that file, and the reveal lets you check it afterwards.