Break an AI answer into specific claims and independently open original sources. Check publisher, dates, scope, and whether the actual passage supports the claim. A working link or agreement from another model is insufficient. Leave unresolved points open and use official channels before consequential action.
Last updated: 2026-10-12
An AI answer says an official website confirms a fact and provides a working link. That can be enough to discourage further checking. Yet a page existing, belonging to the right organization, and supporting a particular claim are separate questions.
This workflow supports everyday research such as opening hours, specifications, and event conditions. For insurance, finance, medicine, or legal consequences, a verification table helps identify questions but does not replace controlling documents or appropriate professional judgment. NIST’s generative AI profile discusses reliability risks including erroneous generation.
Split the answer into small claims
Instead of asking whether an entire paragraph is correct, extract specific amounts, dates, people, locations, and conditions. Separate factual descriptions, inferences, and proposed actions.
For example, “The museum opens Mondays, children enter free, so you can just go” raises at least three questions: whether it opens on the specific date, which ages qualify, and whether a reservation is required. The first two being true would not establish the third.
Copy this verification prompt
1 | Break the answer below into independently checkable claims. |
The result is a queue for review. A model’s “verified” label still requires checking; the generation process should not certify itself.
Ask four questions of the source
First establish whether the actual provider, regulator, or original publisher maintains the page. Then inspect publication date, effective date, and the period described. Last year’s announcement might apply only to last year’s event.
Next check location, product version, and eligible population. Finally locate the passage supporting the claim and inspect exceptions, footnotes, and later revisions. Search snippets can locate a page while omitting those conditions.
| Review result | Record and response |
|---|---|
| Supported | Passage, conditions, review date |
| Partly supported | Exact supported part and remaining question |
| Conflicts with source | Corrected claim and original evidence |
| Unconfirmed | Explicit unknown, without a definite action conclusion |
A company blog may explain its own process without establishing a rule for every provider. An official document can also be authentic but superseded.
Practice on a hypothetical visit
Suppose AI says that arriving before 4 p.m. Monday gives every child under twelve free entry without booking. Your checks show Monday opening on the hours page, “under six” on the ticket page, and no clear booking answer.
Record separate outcomes: opening information is supported, the age statement conflicts with the source, and booking is unresolved. Correct the age and confirm the particular visit date and reservation requirement through the official ticket channel. Labeling the entire answer “mostly correct” hides the unresolved action.
Verify payment and login demands separately
When a message demands urgent payment, login, or a verification code, do not click first and investigate afterward. FTC’s scam guidance recommends contacting organizations through independently obtained trusted channels. AI may organize suspicious requests but cannot guarantee a recipient’s identity.
Retain the version and date behind the final decision. Apply this workflow to AI trip planning and product comparison. Reusing the table is helpful; reusing an old factual conclusion without checking applicability is not.
Sources and review
Sources reviewed:
Illustrative organizational workflows, not product benchmarks or professional advice. Verify AI output, contracts, deadlines, and consequential facts against original sources before acting.
Frequently asked questions
Is an answer trustworthy because it links to an official website?
Not automatically. The site may be authentic while the linked page fails to support the stated amount, date, or eligibility conclusion.
Does agreement from another AI count as verification?
No. Models can repeat the same error. Inspect original documents and relevant official channels instead.
