THE SHORT ANSWER
Separate the response into checkable claims, locate the original evidence, confirm that it supports each claim in context, and record what remains uncertain. An AI-generated citation is a lead to investigate, not proof by itself.
First, decide what needs checking
A fictional character name in a creative exercise needs a different check from a scholarship deadline. Start with claims that would change what you do: dates, eligibility rules, numbers, attributed quotations and statements about cause and effect.
For high-stakes health, legal or financial decisions, this general process is not a substitute for qualified advice and authoritative current information. Match the depth of checking to the consequence of being wrong.
A five-step evidence check
- Extract the claim. Rewrite it as one proposition, keeping its date, location and conditions.
- Find the original source. Prefer the institution issuing a rule, the original paper or the primary document over a summary of a summary.
- Investigate the source. Who produced it, for what purpose, and with what expertise or evidence? Look beyond its own description of itself.
- Match the claim to the evidence. Read the relevant passage and surrounding context. Check whether the source says exactly what the answer implies.
- Record the outcome. Supported, contradicted or unresolved—and why. Keep the source link and date checked.
The third step reflects lateral reading, a practice taught in the Digital Inquiry Group's Civic Online Reasoning materials. Looking outside a website can help evaluate it before investing trust in its presentation.
Evidence & context: Digital Inquiry Group
Worked example: a scholarship deadline
Imagine an AI answer says an application closes on 30 June and is open to all first-year students. This is a fictional example. Separate the deadline from the eligibility claim. Open the issuing institution's current application page, not merely the link's search-result summary.
Check the year, programme, time zone and any exceptions. If an older announcement says 30 June but the current form says 15 June, the answer is not safe to use. If two official pages disagree, contact the issuer rather than averaging the dates or choosing the one you prefer.
| Claim | Evidence to retain | Outcome |
|---|---|---|
| Application deadline | Current issuing page, year and exact deadline wording | Supported / contradicted / unresolved |
| Who can apply | Eligibility section and relevant exceptions | Supported / contradicted / unresolved |
| An attributed quote | Original text or recording and surrounding context | Exact / paraphrase / not located |
Know when to leave a claim unresolved
Repeated agreement from generated answers is not independent evidence. Nor does a long reference list guarantee that the references exist or support the text. NIST's treatment of confabulation is a reminder to inspect outputs rather than infer reliability from their style.
If you cannot locate a source, say so. Remove the claim, qualify it or delay the decision. An honest gap is more useful than a citation added merely to make an answer look finished.
For the underlying mechanism, read why AI hallucinates. For the wider habit of mind, explore critical thinking.
Evidence & context: NIST
Sources & further reading
- Civic Online Reasoning
Digital Inquiry Group. Source-evaluation teaching resources, including lateral reading: investigating a source beyond its own website.
- Generative Artificial Intelligence Profile (NIST AI 600-1)
NIST. Risk-management guidance, including confabulation. It does not establish a universal error rate.
Examples and exercises are illustrative unless attributed to a source. No independent expert review is claimed.
A correction, a counterexample or an experience worth sharing?
Join the conversation ↗