Chase the claim outside the chat window
Asking the model to verify itself is not verification. Teach students to leave the chat and find independent coverage of the claim.
Nov 18, 2026 · 1 min read
The instinctive check on a suspicious AI claim is to ask the same model "are you sure?", which is asking a witness to cross-examine themselves. Teach the exit move instead: take the claim, leave the chat, and hunt for independent coverage in sources that owe the model nothing. Breakstone et al. (2021) found that most secondary students cannot reliably evaluate online sources even when prompted; unprompted, they will not do it for AI output either. (Manning's source-credibility-evaluation-protocol skill adapts fact-checkers' lateral reading into a step-by-step classroom protocol.)
In practice
- Drill the sequence: stop, copy the claim, open a new tab, find two independent sources that cover it, trace any statistic to its origin.
- Checking the AI's own citations is a different job (see "Teach lateral reading on AI citations"); this move works even when it cites nothing at all.
- Set a two-minute budget; fact-checkers work fast because they search about the claim rather than reading deeper into the text.
- Count a claim confirmed only when the confirming source did not come from inside the chat itself.
Quick takeaway: verification happens outside the chat window or it has not happened.