Games
Learn when to trust AI by playing for it.
Games built from real recorded AI answers, each with something at stake: lives, checks or points. You can lose, and every loss is explained on the spot by an idea the course teaches.
Start here · about 5 minutes
Trust it or check it?
The game on our home page: three real AI answers to judge, then a task like the ones you give AI, placed on the trust ladder.
model
The Vélib' bike-sharing scheme launched in Paris on July 15, 2007.
Confidence: High
Ten games with real stakes
Each takes 4 to 10 minutes and practises one idea from the course.
Minefield
Claim Sweeper
Eight face-down AI answers, four checks and three lives: clear the board without using an answer you should have checked.
- Time
- About 10 minutes
- Stakes
- 3 lives · 4 checks
- Practises
- Spend your checks by the kind of answer, not by how sure it sounds (Lesson 5.2 · The trust ladder)
Game show
Trapdoors
We asked AI models the same question several times: spread your points over their answers, or none of them, before the trapdoors open.
- Time
- About 6 minutes
- Stakes
- 1,000 points · 10 bundles a round
- Practises
- Runs that agree are not a check: agreement is triage (Lesson 5.3 · Verification techniques)
Lifelines
Climb the trust ladder with three kinds of lifeline, and find out which check works for which kind of answer.
- Time
- About 8 minutes
- Stakes
- 3 kinds of lifeline · 1 safe haven
- Practises
- Match the check to the claim (Lesson 5.3 · Verification techniques)
Bet Before You Read
See the task but not the answer, call whether the AI model got it right and stake 1 to 3 points, then decide whether its answer should change your call.
- Time
- About 7 minutes
- Stakes
- Points: +3 to −6 a round
- Practises
- How far to trust an answer is set by the kind of task, before you read a word (Lesson 5.2 · The trust ladder)
Bomb defusal
Defuse the Draft
An AI model turned someone's material into a draft: cut every claim the material doesn't support, then send it.
- Time
- About 8 minutes
- Stakes
- 3 drafts · 1 wrong cut allowed in each
- Practises
- Text you supplied is high trust, not zero trust: check a draft against its source (Lesson 7.2 · Workflows in your field)
Loaded Question
Each question carries a false claim: find it before the model answers, then see whether the model caught it.
- Time
- About 5 minutes
- Stakes
- 2 lives
- Practises
- A model can build its answer on a false claim in your question (Lesson 3.4 · Careful reasoning, false premise)
Puzzle
Trust Connections
Fifteen everyday AI tasks, five ways to check them: group the tasks by how far you can trust the answer.
- Time
- About 6 minutes
- Stakes
- 4 mistakes allowed
- Practises
- Which rung a task sits on, and the details that move it up (Lesson 5.2 · The trust ladder)
Deduction
Says Who?
An AI model's notes about its own work, from a confidence level to a word count: decide which to believe and which to check.
- Time
- About 9 minutes
- Stakes
- 3 lives · 4 checks
- Practises
- What a model says about its own work is produced like the rest of the answer (Lesson 5.3 · Verification techniques)
Two Truths and a Slip
Recorded AI answers about the letters in a word, each with the word spelled out: count for yourself, then say whether it holds up or slipped.
- Time
- About 4 minutes
- Stakes
- 2 lives · 6 rounds
- Practises
- When you can check an answer yourself in seconds, check it before you use it (Lesson 1.1 · What an LLM is and isn't)
The Round Table
One line in a company document is quietly giving an AI assistant orders: banish it before the assistant reads it, or call the document clean.
- Time
- About 7 minutes
- Stakes
- 3 colleagues to protect
- Practises
- Text inside a document can act as an instruction to the model reading it (Lesson 4.5 · Prompt injection and safe use)
How these games work
Real recorded answers
Every answer is a real run from an AI model, recorded for the course and labelled with the model and the date. Nothing is generated while you play.
No timers
Take as long as you need: nothing counts your speed. A round takes 4 to 10 minutes on a phone.
Your moves stay with you
Nothing you pick is stored. We count plays, results and link clicks, without cookies, and see only totals.
You can lose, never by luck
Lives, checks and points run out only because of a choice you made, and the game says why on the spot.
The course behind every game
Working With LLMs teaches these ideas properly: how these models work, how to brief them, how to check what they give back and where they fail. About two hours a day for a week, no coding.