Make the AI argue against itself

Models agree too easily, with you and with themselves. A useful classroom counter-move: after any substantive AI answer, require the follow-up "now make the strongest case that what you just said is wrong or incomplete." Where the self-critique is weak, the original was probably solid; where it lands a real punch, the student has found the claim that needed checking, and practised the adversarial reading that transfers to every source. (Manning's ai-expertise-interrogation-designer skill builds full routines on this.)

In practice

• Make it a standing pair of prompts for research tasks: answer, then counter-case, then the student writes two sentences adjudicating.
• The adjudication is the assessable part: which points of the counter-case stick, and how would you check?
• Anything the counter-case flags as factual and checkable goes through lateral reading (see "Teach lateral reading on AI citations").