Teach with the record

Who it is for

The kit can be used in a university course in security, safety engineering, risk, law, or public policy, or with a team or a board deciding how far to let an AI system act on its own. It requires no technical background, only a screen for the practice tool and the self-check.

A one-hour session

1. 10 minutes

   Introduction

   Explain the difference between why a system acted and what let its action go through. Start with the Replit entry, where a code freeze existed only as an instruction to the agent. Then use the problem page to show where prompts and human vigilance sit in the hierarchy of controls from occupational safety.

2. 20 minutes

   Finding the gaps in pairs

   Each pair takes the same incident in the practice tool and decides which gaps it shows, without looking at the record. The Air Canada chatbot is a straightforward first case; the OpenAI and Hugging Face incident is more difficult and includes four close calls.

3. 15 minutes

   Comparing answers

   Pairs compare their answers with each other and then with the record, reading the evidence for each difference. The method notes on the record page explain how each close call was decided.

4. 15 minutes

   Applying the self-check

   Each person picks one AI system they know and one action it can take, and answers the 13 questions in the self-check. Finish by discussing the gaps that came up most often and who could close them.

Poster of the 13 gaps

Display the poster during the session or hand it out as a reference sheet for finding the gaps.

Poster as PDF (A3)Poster as PNG

Exercises

1. Systems that held credentials. In the entries file, find the incidents marked with direct access (G2). What kinds of system are they, and what would it take to remove the credential from each one?
2. Tests that reached outside systems. Read the six entries from the testing, training or evaluation of AI models in 2026 that the record’s findings list. Which gaps appear in all six? For one of them, what would a check at the edge of the test have needed to see to stop the traffic, and what might it still have missed?
3. What the reviewer saw. Pick two incidents marked G7. What did the person who approved the output see, and what would they have needed to catch the error?
4. Specifications. Choose one of the runtime specifications and one agent incident. If the specification had been fully in place, which gaps would have remained?
5. Write the rule. For the Chevrolet chatbot or the Replit agent, write in plain words a rule that a check outside the AI could apply to the action that did the harm. Then say what the rule would stop that it should not, who may change it, and how anyone would know it had changed.

Questions for discussion

• When is a person’s approval a genuine check, and when is it a formality?
• Should an AI system hold credentials at all? What would be needed to remove them?
• Who should be able to stop an AI system, how quickly, and without whose cooperation?
• If an AI system’s account of what it did turns out to be wrong, what record would you trust instead?
• In 2026, tests of AI models reached companies that were not part of the tests. What should count as inside a test?

Feedback and contributions

If you teach with these materials, tell us what worked and what did not, and send anything you made that others could use. With your permission, we will link to it here.