Why is a plausible answer not enough?
A plausible answer is not enough because generative AI can produce a false statement, a fabricated citation or a correct fact used in the wrong context, all in the same confident tone. Fluency makes those problems easier to overlook. The NIST Generative AI Profile identifies confabulation as a risk and lists reviewing and verifying sources and citations in AI output among its suggested actions.
NIST also notes that AI output may include confabulated logic or citations that appear to justify the answer, which can lead people to trust it inappropriately. A neat chain of reasoning is not evidence that the conclusion is right. The guide to AI hallucinations explains why this happens; this guide is about the checking routine.
Which claims should you check first?
Check the claims with the highest consequence if wrong, and the claims most likely to be invented, first. You rarely need to verify every word of a low-stakes internal draft, but you always need to verify the figure in a client proposal. The table is a practical triage for common claim types. Adapt it to your team's work and approval rules.
| Claim type | How to check it | Who decides it is fit to use |
|---|---|---|
| Figures, prices, dates | Recompute or trace to the original record | The author, plus a second person if it goes outside the company |
| Quotations and citations | Open the source and find the exact passage | The author |
| Names, titles, organisations | Check an official or first-party page | The author |
| Statements of cause | Ask what evidence would show it; remove if none | The author and the person who owns the decision |
| Legal, financial, safety or employment points | Primary source plus a qualified reviewer | A qualified person, not the author alone |
How should you inspect a citation?
Open the actual source and find the passage that supports the statement. Check the author or publisher, the date, the scope and any qualification. A real link is not enough if the page discusses a different population, time period or question. Record the source so another person can repeat the check.
If you cannot find the supporting passage, do not ask the model to confirm its own citation and stop there. Ask for the missing evidence, remove the statement or mark it as unresolved. A second model can suggest questions, but agreement between two models is not independent verification. Both may have learned the same plausible error.
How do you check the figures and reasoning?
Recompute important numbers from the supplied data using an independent method, such as a formula, a pivot table or a calculator. Confirm the units, period and comparison base. A total can be right while an explanation of why it changed is unsupported, so separate the arithmetic from any claim about cause.
In a fictional table, revenue moves from 80 to 92 units and costs from 50 to 61 units. The changes are 12 and 11 units, so margin rises from 30 to 31 units. The table alone does not explain whether a new customer, a price change or timing caused the movement. A responsible summary states the change and leaves the cause open, or names the evidence that would settle it. Watch, too, for percent and percentage points being mixed up, which is common in generated summaries.
Want to see how this connects to a role?
Explore the planned team trial ↗How do you fact check an AI summary of a document?
Check an AI summary for what it says and for what it leaves out. Start by taking three or four statements from the summary and finding each in the original. Then read the original's headings or first lines of each section and ask whether anything important is missing. Finally, look for dropped qualifiers: may becoming will, in some cases becoming always, subject to approval disappearing altogether.
If the summary will drive a decision, ask the tool to produce quotes for each point before it summarises. Anthropic's documentation for Claude recommends that pattern for long documents, along with asking the tool to remove any claim it cannot support with a quote. It makes the summary faster to check, though you still need to read the quotes against the page.
What does a full fact check look like?
Here is a deliberately flawed draft for practice. A fictional agency, Kestrel Studio, asked an assistant to summarise a client's supplied product brief into a paragraph for a pitch deck. The brief says the client's planner app launched in March 2025, has a free tier and a paid tier, and that a survey of 200 of its own users found most used it weekly.
AI draft: Launched in early 2024, the app has quickly become one of the most popular planners in the UK, with a recent independent survey showing 85 percent of users open it every day. It offers a free tier and a premium tier.
The check, claim by claim. Launch date: the brief says March 2025, so early 2024 is wrong. One of the most popular planners in the UK: nothing in the brief supports it, so remove it. Recent independent survey: the brief describes the client's own survey of its users, so independent is false. 85 percent every day: the brief says most used it weekly, so both the figure and the frequency are invented. Free and paid tiers: supported.
Reviewed paragraph: Launched in March 2025, the app offers a free tier and a paid tier. In the client's own survey of 200 users, most said they used it weekly. The review note lists four corrections and one removed claim, and the account lead signs off before the deck goes to the client.
Can AI help you fact check its own output?
AI can help organise a fact check but cannot complete one. Useful requests include: list every factual claim in this draft as a numbered list; for each claim, quote the sentence in the attached source that supports it or write Unsupported; and identify any figure that is not in the source. Those turn a paragraph into a checklist.
What AI cannot do is be the independent check. Asking the same tool whether its answer is correct tends to produce reassurance. Running the same question several times and comparing the answers, a technique Anthropic's documentation describes, can reveal inconsistencies, but consistent answers can still be consistently wrong. The final step is always opening the source or recomputing the figure.
What should the review trail contain?
Keep the original claim, the source or calculation used to check it, the correction and any question still unresolved. Name the person responsible for deciding whether the output is suitable for its intended use. That record makes the review inspectable without pretending that every risk has gone.
A review trail does not need special software. A short note at the end of the document, or a comment thread, is enough for most internal work. For client deliverables, follow your normal approval process; AI assistance does not change who signs off.
When is this checklist not enough?
This checklist is an operational routine, not a certification of accuracy. It is not enough for legal advice, regulatory filings, medical or safety information, audited financial figures or decisions about an individual's employment. Those need a qualified reviewer and the primary sources, whatever tool produced the draft.
It can also be more than you need. A brainstorm, a list of meeting agenda ideas or a rough first draft that only you will read rarely needs a claim-by-claim check. Match the depth of the check to the consequence of being wrong.
Sources and scope
NIST, AI 600-1: Generative AI Profile, July 2024 ↗
Anthropic, Claude documentation: Reduce hallucinations, accessed 16 October 2026 ↗