Close Calls: Spot the AI Risk

How do you play Close Calls?

Each round shows your objective, a short workplace or everyday situation and three actions. Choose the action that completes the work and handles the risk, then press Check my choice. The game marks your choice and the best move, highlights the decisive clue and shows an illustrative outcome.

1. Read the objective first. The goal is always to get useful work done.
2. Read the situation. One detail usually decides the answer, such as a policy, a sender address or a permission request.
3. Click an action card. You can change your choice until you check it. There is no timer.
4. Read the feedback: what your choice would lead to, why each action works or fails, and the term behind the lesson in one plain sentence.

Every scenario and outcome is written in advance. Outcomes illustrate the lesson. They are not results from real experiments, and no AI is called while you play. People, companies and data in the game are fictional.

How is each choice scored?

Each level has five rounds, so 10 points. Score 9 or 10 for three stars, 7 or 8 for two and 5 or 6 for one. Results list the topics you handled well and the lessons to revisit. The game keeps your best score and shows your first attempt separately from practice retries. Scores are practice results, not a certification or a compliance check.

What AI risks does the game cover?

The ten levels move from personal information and secrets to fraud, prompt injection, permissions, reliable answers, bias, consent and combined incidents. Each level teaches one family of risk through five situations, some at work and some at home.

Play levels 1 to 3 without signup. Levels 4 to 10 are free with your email, which also brings an email when a new game is added. If you already unlocked another game on this site in the same browser, these levels open without asking again. Progress saves in this browser, so clearing browser data or switching browsers or devices can lose it.

How is Close Calls different from Trust Fall?

Trust Fall asks whether an AI request should be approved, checked by a person or rejected. Close Calls asks you to name the specific risk in a task and pick the practical response, such as sharing less data, narrowing a permission, verifying on a known channel or pausing to escalate.

Both games reward getting work done. In Close Calls, a blanket refusal usually scores zero, because the safer move is often a smaller change that lets the task go ahead.

Where did the lessons come from?

The lessons draw on public guidance from NIST, OWASP and CISA, with a source link in the feedback where one supports the point. The scenarios themselves are my own judgment as a data and AI leader, and every one was challenged by a separate review before publishing.

• NIST definition of personally identifiable information and NIST SP 800-188 on de-identification
• OWASP LLM01 Prompt Injection and OWASP LLM06 Excessive Agency
• CISA guidance on recognizing and reporting phishing
• NIST SP 1270 on bias in AI

The game teaches habits, not legal advice. Your organization’s policies and the rules that apply to you decide what is allowed.