Teach them to catch the model being wrong
The most useful AI lesson isn't how to prompt. It's watching the confident machine be confidently wrong.
Jun 11, 2026 · 1 min read
Students calibrate trust in AI by experience, and most of their experience is frictionless success. Engineer the failure. The OECD and European Commission's AI literacy framework puts critical evaluation of AI outputs at the centre of its "Engage with AI" domain (OECD & European Commission, 2026), and nothing builds that muscle like catching a real error.
In practice
- Live-demo a model on a topic the class just mastered, and let them mark its answer with the rubric they know. Errors they find themselves stick.
- Set a "fact-check the machine" starter: one AI-generated paragraph on yesterday's content, find what's wrong or unsupported.
- Debrief the kind of failure each time (invented fact, outdated claim, plausible-but-wrong method) so students build a taxonomy, not just an anecdote.
- Resist the cynical version. The goal is calibrated trust: right about most things, checkable about everything.
Quick takeaway: students who have caught a model lying once will check it forever.