How to fact-check an AI answer with a claim-and-source table
Check each important AI claim against its source, separate inference from evidence, resolve conflicting dates and preserve uncertainty before using an answer.
An AI answer is only useful evidence when you can check its important claims. Smooth wording, a few links or a confident tone don’t show that the sources support the conclusion. First check the claims that would change what you do next.
The NIST generative AI profile lists confidently wrong content as a risk and recommends checking sources and citations. The method below is something you do yourself, with any AI assistant, including Mono by Reunified, which is currently on a waitlist. It isn’t an automatic fact-checking feature.
1. Turn the answer into checkable claims
Split a paragraph into statements that can each be true, false or unknown. Start with dates, numbers, who qualifies, compatibility and anything you plan to repeat to someone else. A vague “this is a good option” needs clear criteria before you can check it.
Use this fictional workshop source for practice:
Capacity: 24 participants.
Registration closes on 18 October 2026.
Equipment loans are available on request.An answer might turn that into “24 places remain, registration closes at midnight, and equipment is free.” None of those three additions is in the source. Keep the original text next to the answer so you can see where it drifted.
2. Ask for evidence without assuming it was retrieved
Ask for a table instead of a longer, confident explanation:
List each important factual claim separately.
For each, identify the source and the passage that supports it.
Mark inference and missing information explicitly.
Use only the supplied material unless I authorize additional research.
Do not invent a citation or fill a missing date, quantity or condition.This prompt makes the answer easier to review, but it doesn’t make it reliable. Open the cited sources yourself. A generated link can be wrong, and a real page may not say what the answer claims.
Which tools Mono can use depends on your model and connections. A page named in the answer wasn’t necessarily opened. If your setup can’t fetch sources, paste the relevant text yourself or check the source in a browser.
3. Build a claim-and-source table
For the fictional example, a useful review looks like this:
| Claim | What the source says | Decision |
|---|---|---|
| 24 places remain | Total capacity is 24; nothing about places left | Don’t report places remaining |
| Closing time is midnight | A closing date, with no time or time zone | Confirm the exact cutoff |
| Equipment is free | Loans on request; no price given | Ask about conditions and cost |
Note the source title, the link or document location, the relevant passage and the date you checked. The table can be an ordinary note.
Keep facts from the source apart from your own calculations or conclusions. If you calculate a total, keep the inputs and units so someone else can repeat it. Correct math can still mislead if one input covers a different time period or group of people.
4. Check authority, scope and freshness
Go to whoever is responsible for the fact: the event organizer for dates, the software publisher for supported versions, or the original research for a study’s finding. A search snippet or another AI summary is a lead to follow, not a replacement for the original.
Read the conditions around the matching sentence. A feature that’s only in a beta, one plan or one region shouldn’t become a claim about everyone. The publication date and the event date are different things: an older page can describe a future event, and a recent repost can repeat outdated information.
When sources disagree, note it and look for a clear update or the person responsible. Don’t split the difference between two dates or pick the one that more summaries repeat.
5. Revise the answer around what is known
A supported version of the example is: “The workshop has capacity for 24 participants. Registration closes on 18 October 2026; the exact cutoff and current availability need confirmation. Equipment loans are offered on request, with conditions to check.”
Leave open questions visible instead of covering them with a polished guess. If a fact affects a booking, purchase or publication, ask the right person a specific question. For important decisions, rely on the official source or a qualified expert, not on this checklist alone.
Check the edited answer against the table again. A second model can point out problems, but two models agreeing doesn’t prove the fact.
6. Save a usable evidence trail
Keep the final wording, checked sources, review date and open questions together, for example in a note in Mono’s Vault. Our Markdown notes guide covers ongoing notes, and a conversation export helps when you want to keep the discussion itself.
Check time-sensitive claims again before you reuse them. A source checked for one workshop, software version or date doesn’t automatically apply to the next one. To turn notes into a work plan, the meeting-notes recipe uses the same idea: keep what people actually agreed apart from what the AI guessed.
