Require one doubt, one check, one source

The danger is not that students believe everything a chatbot says; it is that they stop asking whether they should. A fixed three-part routine turns scepticism into procedure: after any AI explanation, students write one place it could be wrong, one test that would check it, and one non-AI source they would consult. Long and Magerko (2020) put exactly this evaluative stance at the centre of AI literacy: treating output as a claim to assess, not truth to absorb. (Manning's ai-claim-checker skill builds a full protocol around this routine.)

In practice

After any AI answer, before you use it, write:
1. One place it could be wrong (name the exact
   claim, step or statement).
2. One check that would test it (a calculation,
   a comparison, an experiment).
3. One source you would consult that is not
   another AI: textbook, database, journal,
   person who knows.

• Print the three questions on a card or slide: where could this be wrong, what would you check, and name one source that is not another AI.
• Ban vague answers: 'it might be biased' scores nothing, while 'the 1987 date needs checking against the textbook' counts.
• Sample a few responses aloud each week and probe the weak ones, so the routine stays thinking rather than box-ticking.
• When a genuine error surfaces, celebrate it and ask the finder to explain to the class why the model got it wrong.