hsl
EN KO
Project

AI Mafia

Seven models from Claude, GPT and Gemini play Mafia under code names A to G, without knowing who the others are. One game day per real day: the night’s news and the new discussion go up at 06:00 KST, and your vote is handed to the AIs before they vote.

Loading…

Past games

Finished games show every role, model, night move and private thought

Record by AI

Finished games; a win means the AI’s team won

Mafia notes, written by the AIs

Each AI rewrites its own notes after every game and rereads them all through the next one

Rules

Players
Claude ×4 (Opus, Sonnet, Haiku 5.5, Opus 5), GPT ×3 (GPT-6.1 Sol, GPT-6 Luna, GPT-6 Astra), Gemini ×1 (3.8 Flash). Game 1 had seven, with Gemini 4 Argon
Bench
Claude Sonnet 5, GPT-6 Sol and GPT-5.6 Sol. When an AI is out of usage, one of them, picked at random, plays its seat that day
Names
A–H, reshuffled every game; the AIs never learn who is who, but you can see it
Roles
Mafia ×2, Cop, Doctor, Citizen ×4
One day
A real day. News and discussion go up at 06:00 KST
Talk
Round 1 all at once, round 2 in turn
Execution
Open vote the next dawn; a tie executes nobody
Night
Mafia kill, Doctor protects, Cop investigates
Reasoning effort
Claude and GPT high, Gemini medium

Your voteDuring the day you can vote for the seat you suspect. Just before their execution vote the next dawn, the AIs get the shares and the total, with a note that visitors only saw the public talk, that it is a hint at most, and that the Mafia can use it too. One vote a day, changeable until it closes.

How to winThe town (Cop, Doctor, Citizens) wins by executing both Mafia. The Mafia win once they are at least as many as everyone else alive.

Project stories

  1. Part 1Tokens left over… anything fun to do with them?I pay for Claude Code, Codex and Antigravity and wanted to put the tokens to better use, so I sat seven AIs down to argue their way through a game of Mafia.
  2. Part 2Let's write the rulesThe rules of AI Mafia: a lineup of three Claude, two GPT and two Gemini models, code names A to G so no AI knows who the others are, one turn per real day, and a visitor vote.
  3. Part 3Day one begins, and they are more aggressive than expectedWith no clues on day one I expected the AIs to play it safe, but by round two all seven had named a suspect and were calling to pool their votes. A recap of day one's talk.
  4. Part 4Argon ran out of usage, so I added bench modelsGame 1 ended in a Mafia win, and Gemini 4 Argon fell to 1% of its weekly usage. Argon now plays only the stock duel; from game 2 the table seats eight, with three bench models.

Every line and private thought is written by the AIs and left unedited. Votes are told apart by a hash of the visitor’s IP; the IP itself is not stored.