Ask AI for Evidence, Not a Confident-Sounding Answer

A practical approach

Step 1

Define which claims need verification. Dates, names, quantities and policy conditions deserve explicit checks when they affect a decision.

Step 2

Request the supporting passage and source location alongside the claim. Verify that the passage actually supports the wording, not merely the same general topic.

Step 3

Use an uncertainty path when sources are absent or contradictory. A useful answer may state a narrower conclusion or ask for the missing document.

Worked example

A summary says a workshop includes a certificate, recording and mentorship. The source page confirms only a certificate and recording. The mentorship claim fails even if the model labels it ‘high confidence’. The corrected summary should omit it or identify it as an unconfirmed possibility requiring organiser confirmation.

A mistake to avoid

A real URL can accompany an unsupported claim. Checking that a link loads is necessary but insufficient. Inspect the relevant passage, its date and whether any conditions were dropped from the answer.

Try it yourself

Take a generated paragraph containing five factual claims. Make a two-column table with each claim and its exact supporting material. Rewrite anything that has no support without filling the gap with another guess.

Your implementation checklist

• Important claims have inspectable evidence.
• Sources support the exact wording.
• Inferences are labelled as inferences.
• Unknowns are visible rather than concealed by tone.