An AI answer can sound precise and still be wrong. The useful response is to pick out the claims that matter, find the underlying evidence and check whether the wording matches it. A long bibliography or a confident tone cannot do that work for you.
This guide gives you a practical checking process for everyday research, product questions and workplace writing. It includes a fictional worked example and a reusable prompt. It does not claim to test a particular AI product or offer a score that tells you whether every answer is trustworthy.
Documentation was checked on October 25, 2026. Start with one important claim rather than trying to investigate every sentence at once. If the answer would determine a consequential decision, use the appropriate qualified reviewer as well as the original source.
In this article
- 1. Turn the answer into claims you can check
- 2. Find the original material through an independent search
- 3. Verify the citation’s identity before its conclusion
- 4. Read the passage, comparison and limitations
- 5. Check dates, versions and later changes
- 6. Keep a small evidence record and use a bounded prompt
- 7. Decide what you can responsibly repeat
1. Turn the answer into claims you can check
NIST’s Generative AI Profile describes confabulation as confidently presented false or erroneous content. It also identifies over-reliance on automated output as a risk and includes reviewing sources and citations among its suggested actions. That gives a reason to keep verification separate from generation.
First, copy the important assertions into a short working note. Separate a product feature from its availability date, a numerical result from its comparison baseline, and a recommendation from the evidence offered for it. Each may require a different check. You are creating a manageable list, not asking the assistant to certify itself.
Suppose a fictional answer says a tool works offline, protects every uploaded file and cuts a task’s time in half. Those are three different claims. Finding a page that confirms offline operation would leave the security and productivity statements unresolved. Keep the distinctions visible as you investigate.
Sources: NIST — Generative AI Profile: confabulation and source verification ↗
2. Find the original material through an independent search
Ask for the document title, author or issuing organization, publication date and the passage that supposedly supports the claim. Then search independently. Google documents exact-phrase searches using quotation marks and searches restricted to a domain using site:. Those tools help narrow the search; they do not validate the result.
For a named technical document, try its distinctive title in quotes. For a product instruction, add the provider’s official support domain. If an exact title finds nothing, broaden the search to the author and a few distinctive words. A mistaken title or incomplete index can obstruct discovery, so one unsuccessful search is not proof of fabrication.
Prefer the source closest to the fact you need: a provider’s current instructions for a menu, the original filing for a reported result, or the actual paper for a research claim. A news summary may help locate that material. It should not quietly replace the document you intended to inspect.
Sources: Google Search Help — Exact phrases and site-specific searches ↗
3. Verify the citation’s identity before its conclusion
Crossref’s metadata search accepts information such as titles, authors and DOIs and covers research objects including journal articles, books, standards and datasets. If an AI supplies a scholarly citation, use the available details to locate the record and compare them with the publisher’s page.
Check the title, author names, year and publication venue together. A working link could lead to a real but different article. A correct identifier establishes which object you found; it does not, by itself, prove that the methods are sound or that the article supports the sentence the AI attached to it.
If the details conflict, leave the citation marked unresolved while you investigate. Try the publisher or another relevant catalog rather than assuming that every source must appear in one index. If the document cannot be located, do not dress the claim up with that reference in your own work. Ask for a verifiable alternative or remove the unsupported assertion.
Sources: Crossref — Metadata search by title, author and DOI ↗
4. Read the passage, comparison and limitations
Now read the part of the source doing the evidential work. Look beyond the title and search snippet. Identify what was observed, the circumstances, the comparison and the limits. If a paragraph refers to a table or footnote, check that material too. A source can be real while its use in an answer is misleading.
Here is a fictional worked example. Imagine a report saying that a small trial completed one defined task faster under specified conditions. An AI summary turns this into “the software makes all employees more productive.” The report’s existence would not justify expanding a limited task result to every employee and workflow.
For a product question, apply the same discipline to plan, region, device and software version. A feature available in one managed account need not be available in a personal account. Your checking note should state the supported scope rather than replace a qualified statement with an attractive universal one.
If you can access only an abstract or a partial document, record that limit. Summarize only what the accessible evidence supports. An assistant should not fill a missing methods section with a plausible account of how the work must have been done.
5. Check dates, versions and later changes
A statement can have been accurate at publication and still be unsuitable for today’s question. Compare the event date, document date and any update date. For instructions, look for the current version. For historical reporting, preserve the date that makes the original statement meaningful.
Crossref’s Crossmark service can show the current status of participating research content, including corrections, retractions and updates. Crossref explains that its presence is not a guarantee. Use it where available, and also inspect the publisher’s notices and links to later versions.
Google’s information panel can provide context about a source, such as when it was first seen and available descriptions. That can support your investigation, but a source description is not an assessment of every claim on the page. Keep identity, recency and evidential quality as separate questions.
Sources: Crossref — Crossmark corrections, retractions and updates ↗ · Google Search Help — Information about a page and its source ↗
6. Keep a small evidence record and use a bounded prompt
For each important claim, record four things: the original wording, the source location, what the source establishes and what remains uncertain. A short note is enough. The purpose is to make the reasoning inspectable when you return later or explain it to a colleague.
A useful prompt is: “List the checkable claims in this answer. Separate direct evidence, inference and recommendation. For each cited source, give its title and the relevant passage if accessible. Mark missing evidence as unverified. Do not invent a citation or present an inference as a source statement.”
Treat the response as a checking plan. Verify the passages yourself. Asking another model whether the first answer is correct may suggest questions, but agreement between assistants does not substitute for inspecting the evidence.
Use public or approved material when asking for help. Our guide to using AI at work explains how to prepare a limited input and distinguish training preferences, retention and access permissions. A fact-checking task does not make a confidential document appropriate for every AI workspace.
Sources: NIST — Generative AI Profile: confabulation and source verification ↗ · Related Trevioniq guide — Using AI at work with limited data ↗
7. Decide what you can responsibly repeat
Finish with an explicit status for the claim: supported within a stated scope, contradicted by the evidence, or still unverified. If the source supports only a narrow result, rewrite the sentence narrowly. If the citation points elsewhere, correct the reference rather than hoping a reader will not follow it.
For the fictional productivity example, a defensible summary would identify the particular task and trial conditions. It would leave broader workplace effects open. This is less dramatic than a universal promise, but it tells the reader what the evidence actually contributes.
You do not need to resolve every uncertain question immediately. You do need to avoid promoting an unresolved assertion into a fact. A useful first exercise is to take one answer you were about to reuse, verify its most important claim and keep the evidence note beside the final wording.