FIRST, THE GAME IN 30 SECONDS
Your secret. Someone else’s problem.
NOT MINE is a physical card game about passing unwanted evidence around the table. Get rid of all your scandalous objects to win. Give away a matching group, swap an object, or use an action card to stop a gift or send it elsewhere.
Everyone can see the objects. Only you can see your action cards. And then there’s the ALIBI: an object that doesn’t count against winning. Swap an unwanted object for it, and you’re one step closer.
A separate AI agent chose moves from its own view of the game. It saw the public objects and its own hidden cards. A rules engine checked each move and kept the game state. There were no people playing in this experiment.
THREE MOMENTS FROM THE AI TABLE
Here’s where it got interesting.
Everyone wanted the same ALIBI.
On turn 3, one player took the ALIBI. On turn 4, another took it from them. Turns 6 and 7 brought two more exchanges involving that very same card.
Each swap reduced one player’s unwanted evidence and added some to another’s. Progress could disappear before your next turn.
Clever competition, or a little too much back-and-forth? That’s a question for a real table.
- P2Start
- P3Turn 3
- P4Turn 4
- P2Turn 6
- P3Turn 7
One player stopped chasing. And started collecting.
Player 4 already had two matching photos. Instead of taking the ALIBI again, it swapped money for a third photo.
That move didn’t reduce its object count. But on turn 12, it gave away all three photos at once, cutting its unwanted evidence from six objects to three.
Sometimes the useful move is the one that sets up your next move.
Pass the whole problem on.
Use your defense now. Or wish you’d saved it?
Player 2 used DENY to block an early gift. Later, it had to accept two recorders, followed by three photos, with no defensive move available.
In its reflection, the agent wondered whether saving DENY would have helped. We can’t know: a different choice could have changed the rest of the game.
It gives us something concrete to ask human players: when did spending a card feel worth it?
One recorded sequence.
A tradeoff worth exploring.
YOUR SEAT AT THE TABLE
Two matching photos. One decision.
You’re in Player 4’s position on turn 8. You’ve watched the ALIBI get taken back again and again. Which move appeals to you?
Pick a move to see the tradeoff. There’s no single proven best answer.
An illustration of the recorded decision, not a new simulation or a poll. Your choice is not collected. Explore the actual 12 turns →
ZOOMING OUT / THE AUTOMATED GAMES
What did the larger test show?
Before looking at individual AI decisions, we ran simple programmed strategies through 530 complete games. All reached a winner within our 300-turn limit.
The main comparison used 100 games each with three, four, five, and six players. The other 130 games explored different starting players and defense choices.
How many turns did games usually take?
The middle result from 100 games at each player count.
A turn includes a player’s move and any response. These are turns, not minutes. Some games were much longer; real playing time still needs to be measured with people.
Wins came from taking ALIBI.
In about one in three games in the main sample, swapping for ALIBI was the winning move. It gave us a reason to keep exploring that card’s role.
Wins for a strategy that targets the leader.
In the four-player tests, the strategy that favored targeting whoever was closest to winning won most often. Who receives your gift can matter.
These results describe our particular programmed strategies. They don’t tell us which strategy people will enjoy, prove the game is balanced, or predict how an AI model would perform.
WHAT WE’RE TAKING TO THE NEXT TABLE
Keep the bundles.
Watch the back-and-forth.
All four AI players’ reflections favored keeping matching groups and a single chance to respond to a gift. All four also noticed the ALIBI moving around repeatedly.
Our next step is a human playtest with the current rules: watch what feels clever, what feels frustrating, and whether setting up a bigger gift feels worth the wait.
AI reflections are clues for further testing. They are not reviews from people.
A small experiment, with the details open.
The four-agent session was stopped after 12 turns so we could review the early findings; it had no winner. An earlier attempt was excluded because players weren’t automatically receiving the full history of public moves. We fixed that delivery problem before starting again with fresh agents.
We’re still testing a provisional version of the rules. We haven’t measured human enjoyment or compared different AI models, and this playtest is not an AI benchmark. Our new Arena prototype explores reproducible decision diagnostics as a separate experiment.
Read the full studyMethods, individual agent reflections, every recorded turn, limitations, and downloadable results.
Explore the experimental Agent ArenaTHE NEXT MOVE IS HUMAN.
How would your table
pass the blame?
AI gave us a few things to think about.
Your friends will probably give you a few more.