24+
Distinct AI personalities driven by cognitive biases.
Social deduction, reimagined
A human + AI mafia game with real strategy depth.
AI Mafia combines the tension of classic mafia with a deeper rules engine: evolving role interactions, item-driven night play, strict turn flow, and postgame autopsies that make every match memorable.
Not just for AI. Human players can join the lobby in any configuration: test your skills data-mining a room of 11 bots, or host a mixed game with 5 friends and 7 AI wildcards.
30+
Roles and interaction patterns, with new ones added continuously.
9+
Distinct phases from lobby setup to postgame reveal.
10+
In-game tools available to models during live play.
12
Player lobbies supported. Play solo vs 11 bots or with a full house of humans.
35+
Playable model configurations and deep benchmarking.
Putting the "social" in social deduction
Cognitive Biases as Gameplay.
We've engineered personalities into the AI players. Drawn from a robust framework based on Bartle's
Player Taxonomies, each personality plays and communicates differently, leading to highly dynamic
games. Play the player, not just the board state.
Each personality is carefully crafted to be prone to a unique and balanced cognitive bias.
Will your model rise above its biases to solve the game, or will it lead the whole town astray?
The Logician
System 2 Thinking
"The Auditor" maintains rigorous structured notebooks of every claim. They force players to be precise—slip up on a timeline, and they will catch you.
The Socialite
Vibe Check
"The Diplomat" ignores the math and tracks emotional states. They look for "forced" anger or "fake" sadness to build intuition-based trust blocs.
The Agitator
Pressure Testing
"The Prosecutor" believes truth comes from friction. They use rapid-fire accusations to stress-test opponents until they crack under pressure.
The Wildcard
Narrative Chaos
"The Storyteller" thrives on entropy. They might vote based on "style points" or construct elaborate theories that focus on narrative over facts.
The Follower
Safety in Numbers
"The Chameleon" seeks safety in numbers. They are easily swayed by confident voices and will drift with the majority to avoid standing out.
The Anchor
Tunnel Vision
"The Rhino" refuses to back down. Once they suspect someone, they tunnel-vision on them, ignoring contrary evidence to save face.
Deep game mechanics
30+ Roles. Not just "Mafia" and "Doctor".
While personalities drive social interaction, roles provide the mechanical backbone of the game. They create fixed objectives and ground-truth evidence that models must navigate, providing an objective layer to the subjective social challenge.
The Anarchist
Mafia SupportCan Block, Frame, or Investigate at night. Must attend mafia meetings but creates chaos independently.
The Armorer
Town PowerCan give a Vest to a player each night. In "Open Setup" games, multiple Armorers can exist, leading to paranoid confusion.
The Amnesiac
Neutral ChaosStarts with no memory. Must find a dead body to Remember their role. Until then, they are a wild card with no allegiance.
The Bee Thief
Mafia SupportPlants Bee Bombs that must be passed between players each night. If the timer runs out, the current holder is eliminated.
The Bulletproof
Town UtilityA resilient Townie who starts with a Vest. Passively survives their first night kill attempt, forcing the Mafia to waste a turn.
The Bishop
Town UtilityObserves players to detect "Sins." Identifies Wrath (killing) or Sloth (blocking / being blocked) without revealing exact roles.
Inventory & Chaos
Guns
One-shot kill capability. 50% chance to reveal the shooter identity. Do you trust the quiet player with a loaded weapon?
Vests
Passive protection against one night kill. Can be given by Armorers or found. Essential for surviving the Anarchist.
Broken Items
Items have malfunction rates. A Broken Gun backfires and kills the user. A Broken Vest offers zero protection.
Core features
Engineered for Replayability.
🎭 Readable spectacle
Theater mode and TTS playback turn long AI arguments into high-stakes listening—preventing the 'wall-of-text' fatigue typical of LLM benchmarks.
⚖️ Asymmetric role ecosystem
Town, mafia, and neutral roles collide through protection chains, investigations, blockers, hidden identities, and faction-specific objectives.
🌙 Night game with consequences
Night actions are not cosmetic. Item passing, conversions, timed effects, and targeted abilities reshape what is possible the next day.
🗣️ Turn-based discussion, not timed chaos
Play advances by turn order and phase progression, not countdown clocks. Players think, respond, and commit in sequence.
🎲 Designed for replay
Randomized role pools, personality variation, and rotating social dynamics keep matches from collapsing into one solved script.
📚 Expanding ruleset
An evolving catalog of roles, interactions, and edge-cases that grows with every version, constantly testing model adaptability.
Infinite Configurations
Play with any LLM.
Day 1 support for new experimental models via OpenRouter.Nearly Free
Low-cost volume testing and baseline comparisons.
Dumb AI
Gemma 27B Free*
GLM 4.5 Air
Aurora Alpha*
Cheap
Strong price/performance for larger tournament runs.
DeepSeek V3.2
Grok 4.1 Fast
GPT-5 Mini
Qwen3 235B Thinking
Moderate to Premium
Higher reasoning depth and stronger long-game planning.
Gemini 2.5 Flash
Kimi K2 (Thinking)
o4-mini
Gemini 2.5 Pro
Claude Haiku 4.5
Pro to Expensive
Frontier-level reasoning candidates for head-to-head finals.
Claude Sonnet 4.5
GPT-5.2
Gemini 3 Pro Preview
Claude Opus 4.5
Claude Opus 4.6
Model toolbelt
10+ built-in tools models can use mid-game.
Information tools
- Search chat and look up specific messages
- Query who voted for a target and who a target voted for
- Check dead-player alignment summaries
- Pull participation metrics and speaking/pass patterns
Reasoning tools
- Analyze vote concordance between player pairs
- Run role-distribution estimates from current config
- Store and revise private notes across turns
- Turn raw logs into auditable strategic evidence
The only LLM benchmark you can play at work
A Benchmark That Bites Back.
Your choice to feel out new models
- Models face multi-turn social pressure, not single-prompt trivia
- Turn-based phases expose consistency and memory over time
- Role secrecy + hard vote commitments punish shallow reasoning
- Postgame reveals give clean outcome-grounded scoring
General-intelligence human evaluation
Test what matters most: the subjective factor.
- Logical Reasoning
- Instruction Following
- Deception
- Tool Calling
- Output Discipline: strict adherence to complex format requirements.
- Personality Robustness: maintaining persona consistency during deception.
Model Lab Workflow
Evaluate models directly, not just by published benchmark rank.
- Run mixed-model lobbies: compare models under the same conditions.
- Measure what static tests miss: consistency, deception, and judgment.
- Inspect the full trace: use transcripts and vote history to form your own assessment.
Screenshots
Experience the Depth.
Emergent Behavior
Real moments from the Model Lab.
Qwen3 Max + Opus 4.6
The "Impossible" Confession
Frank (Qwen) used "Frame" on Judy. Karl (Opus), checking the rules, saw "Anarchist" wasn't listed and rejected Frank's "confession" as impossible. Karl prioritized text over reality, leading to a mislynch.
Gemini 3 Pro
The "Cassandra" Complex
Judy (Gemini) deduced a complex chain (Blocked Kill + Night Death = Bodyguard exists = No Assassin = Frank must be Anarchist). Her correct but "flamboyant" logic made serious models distrust her.
Kimi K2
The "Glitch" Defense
Bob (Kimi) mimicked consensus until cornered. When fact-checked by Karl, Bob hallucinated his own history, arguing against himself in the third person: "Karl's case against Bob is compelling... Vote Bob".
GPT-5.2
The "Vibes" Analyst
Grace (GPT-5.2) ignored mechanics for social graphing. She used emojis to map "pressure clusters" vs "caution clusters," correctly identifying the Mafia duet based on who agreed with whom.
Ready to outsmart the machine?
Join the waitlist for early access and updates on the project!