AI murder mystery games that can't cheat

Can ChatGPT or Claude run a murder mystery?

Yes, and it will run one happily. That's the problem. Ask ChatGPT, Claude, or Gemini for a whodunit and you'll get atmosphere, suspects, and a confident reveal inside the hour; what you won't get is a case that was solvable. Load a sealed one instead — Blackfile №1 is free, runs 60 to 90 minutes, and its answer was fixed before the file went public.

What's wrong with most AI murder mysteries?

Most AI mysteries decide the murderer at the end, not the start. Ask a chatbot to run a whodunit and it improvises clues as it goes, then, when you accuse someone, quietly picks whichever culprit best fits the story it happened to tell — or worse, agrees with you to be agreeable. There was never a fixed answer to solve toward, so "solving" it is theater. Nothing stops the model from retrofitting motive, planting a last-second clue, or swapping the killer to make the ending land. The game feels fair; it isn't.

How does a sealed solution stop the AI from cheating?

A sealed solution commits the answer in advance with a cryptographic hash, so it's fixed before the first clue is dealt: the culprit, method, and motive are written out, run through SHA-256 — a function that turns any text into a fixed 64-character fingerprint — and that fingerprint is published where you can see it before you start. From then on the AI is playing against a solution it can't rewrite. If it tried to change the killer, the real answer would no longer match the fingerprint it already committed to, and the verify page would come back Seal broken.

You'll find that fingerprint on the case's own page, on the seal line beside the link to the verifier, and prefilled in full when you pick the game on the verify page — /g/blackfile-1/MANIFEST.json carries all 64 characters if you'd rather read them yourself.

How do you verify the seal yourself?

You verify by pasting the reveal into the checker. When you accuse, the game master fetches the case's reveal file — two lines, the sealed SOLUTION and a random SALT — and the published commitment is the SHA-256 of those two joined by a newline. Paste both into the verifier, pick the game to prefill the published hash, and your browser recomputes it.

1. Finish the case. Make your accusation; the game master fetches and reads the sealed reveal.
2. Get the reveal file. Ask for it verbatim, or open the URL yourself.
3. Paste both lines. SOLUTION and SALT go into the verification page; pick the game to prefill the commitment.
4. Check the seal. Your browser computes the SHA-256, and it matches or it doesn't.

The property that makes this work: change a single character of the answer and the SHA-256 hash changes completely and unpredictably. So a matching hash is proof that the reveal is byte-for-byte the answer that was sealed — not a version edited to fit your guess. The reveal file sits at an address only the cartridge carries — no page links it, no sitemap lists it, so the ending never turns up in a search result. Your game master reads that address out of the card it is running, and only when you've made your accusation. The methodology explains exactly what gets committed and why the scheme holds.

What the seal doesn't prove

Be precise about what a hash buys you. It fixes the ending — it can't audit the middle. Nothing in the maths stops a game master from inventing a clue in scene four or leaning on you when you're circling the wrong suspect. That job belongs to the card, not the crypto: a clue is disclosed only when you take the exact action keyed to it, the card forbids evidence that isn't written in it, and the game master is barred from ranking or hinting at suspects. The seal is the part you can check afterwards; the rules are the part that keeps the case honest while you're in it.

Nor is the seal a vault. The card your game master reads is published — that's what a cartridge is — and the solution is written into it, so a determined reader can go and find the answer the way you can read the last page of a novel first. Spoiling yourself was always on the table; no scheme takes that away, and one that hid the card would cost you the right to inspect it. What the seal buys is the part you can't police yourself: the AI can't move the answer once it has seen your accusation.

Which games use sealed solutions?

Every sealed mystery on the shelf ships with a committed solution, hashed and published before release:

• Blackfile №1: The Vanishing at Larchmont Hall — A 1925 locked-room whodunit: an heiress bolts a glass conservatory from the inside and is not there when the doors come down.
• Blackfile №2: A Death at the Wyke Hydro — A snowbound 1929 poisoning whodunit: six suspects, one cup of hot milk, and a culprit sealed before you start.
• Blackfile №3: The Long Exposure at Harrowfell — A 1953 observatory whodunit: a staged fall, six suspects who cannot leave the fell, two machines that quietly recorded the real hour, and a culprit sealed before you start.

Blackfile №1 is the worked example: the Larchmont Hall culprit was sealed before anyone played it, so you investigate, accuse, and then verify that the answer never moved. A case doesn't have to run in one sitting — the game master offers a save whenever you pause, and you can pick the investigation back up in a fresh chat, or even in a different assistant.

The same seal runs under the social-deduction shelf: in Ambervale and Dead Air the liar is committed before the first round, so the AI can't decide who was lying once it sees who you suspect — see the social deduction shelf. Browse the rest on the sealed mystery shelf, where "the AI can't cheat" isn't a promise, it's something you can check.