How to fact-check AI output: a step-by-step process

Fact-checking AI output comes down to three moves: decide which claims matter, verify those against a source you can open, and make the model do part of the work by labeling its own uncertainty. Done as a routine, it takes a few minutes per document and removes most of the risk of using AI at work.

Here is the process, step by step, followed by a checklist you can copy.

Step 1: triage by stakes

Not everything needs the full treatment. Sort the output by what happens to it next.

• Stays with you. A brainstorm, an explanation to help you understand something, a rough outline. Verify only what you will act on.
• Goes to colleagues. Meeting notes, internal summaries, draft plans. Check the specifics (names, dates, numbers, commitments) and label anything unverified.
• Goes outside, gets filed, or drives a decision. Client emails, published content, reports, anything with a number in it, anything legal or financial. Full check of every claim, every source, every figure.

The test that makes triage fast: would you sign your name to every sentence if a skeptical reader challenged it? For the third category the answer must be yes before it ships.

Step 2: separate the claims from the prose

Fluent writing hides claims. A paragraph that reads smoothly may contain six checkable statements, and you will skim past four of them. So extract them first.

Ask the model that produced the draft, or any model, to do the extraction:

List every factual claim in the text below as a numbered list. A factual claim is any statement that could be true or false: a name, number, date, statistic, quote, citation, event, or description of what a company, product, or person did. Do not include opinions, recommendations, or general statements. For each claim, note whether it came from material I provided or from elsewhere.

Text: [PASTE THE DRAFT]

Now you have a list instead of a paragraph. Most work drafts yield five to fifteen items. That is your checking budget.

Step 3: verify each claim by type

Different kinds of claims fail differently, so they need different checks.

Facts and dates

For each, find a primary source: the company's own site, the regulator's page, the original announcement, the record in your system. Secondary summaries, and especially other AI answers, do not count. If the fact could have changed recently (a price, a job title, a rule), the source needs a date on it.

Numbers and statistics

Three questions for every number: Where did it come from? Is it current? Is the arithmetic right?

If the model produced a statistic you did not supply, treat it as invented until you find the survey, report, or dataset it belongs to. Search for the figure together with its claimed source. If you cannot find it in a couple of minutes, cut it or replace it with a number you can source.

If the model did arithmetic (totals, percentages, growth rates), recompute it in a spreadsheet or a calculator. Models handle numbers as text and produce close-but-wrong answers with perfect formatting. Where possible, ask for the formula instead of the result and let real software do the math.

Quotes

Search for the quote verbatim, in quotation marks. A real quote will turn up in the original source. An invented one will turn up nowhere, or in a slightly different form attributed to someone else. If the quote is from a document you supplied, use Ctrl+F on the document itself. If it is not there word for word, it is not a quote.

Citations and links

Open every one. Then check three things: the title, author, and year match what the AI said; the source is what it claims to be (a peer-reviewed article, a government page, a company blog); and the source actually contains the point it is cited for. The third check is where most "real but misused" citations fail: the paper exists, but it says something different.

Links deserve their own warning. A URL can be well-formed and dead, or lead to a page that does not match the description. Never pass on a link you have not clicked.

Code, formulas, and technical output

Run it. A spreadsheet formula either returns the right value on rows where you already know the answer or it does not. A script either passes its tests or it does not. For code, also check that every library, function, and API the model used actually exists in the version you are running; models sometimes invent plausible-sounding function names. Ask the model to write a few test cases alongside the code, then run those too. If you are not the person who can evaluate the code, find that person before it goes anywhere near production. The software developers playbook goes deeper.

Summaries of your own documents

The risk here is addition and omission: a point that was never in the source, or a caveat that was dropped. Spot-check at least three specific claims against the original, choosing the ones with numbers, names, or deadlines. Then ask the model to quote the exact passage that supports each key point, with a page or section reference, and confirm the quotes are real. If the summary drives a decision, read the relevant section of the original yourself.

Step 4: make the model show its uncertainty and sources

Models are surprisingly capable of flagging their own weak spots when you ask explicitly. They will not volunteer it. Build these requests into your prompts.

• Label the provenance of every claim. From provided material, from search, or from memory. Memory is where the risk lives.
• Rate confidence. "Mark each claim high, medium, or low confidence, and explain what would change your mind." Low-confidence items become your checking list.
• Quote, do not paraphrase. "For each claim drawn from the material, include the exact sentence it is based on."
• List what is missing. "What would you need to verify before this could be sent to a client?" This surfaces gaps you did not think to ask about.
• Challenge the premise. "Does my question assume anything you cannot confirm?" This catches leading questions before they produce leading answers.

Here is a single prompt that combines them. Run it on any draft before the manual checks; it usually shrinks the list of things you need to verify by hand.

You are a rigorous editor preparing this text for a skeptical reader. Your job is to expose every unverified claim, not to defend the draft.

Text to review: [PASTE THE DRAFT]

Source material the draft was supposed to use, if any: [PASTE SOURCES OR WRITE "NONE"]

Do the following, in order:
1. List every factual claim as a numbered item.
2. For each, label its origin: SUPPORTED BY SOURCE (quote the exact sentence and its location), FOUND BY SEARCH (give the link), or UNSUPPORTED (nothing in the sources or search confirms it).
3. For each UNSUPPORTED claim, say what a human would need to check and where they would most likely find it.
4. Point out any claim that contradicts the source material.
5. Point out any assumption in the draft that the sources do not confirm.

Constraints: Do not add new facts. Do not soften your findings. If you cannot determine a claim's origin, mark it UNSUPPORTED. Ask me one clarifying question first if the scope of the review is unclear.

One caution: a model can label a claim SUPPORTED and still be wrong about it, because the labeling is itself generated. This step reduces the manual work; it does not replace Step 3 for anything in the high-stakes tier.

Step 5: use grounded tools when facts are the point

Two kinds of tools change the odds in your favor.

Search-grounded assistants put live web results into the model's context before it answers, and show you the sources. Perplexity is built around this. ChatGPT, Claude, Gemini, and Microsoft Copilot all offer search modes; check the current documentation for how to turn them on. Use them for anything current: prices, people, regulations, product features, news.

Source-grounded tools answer only from documents you provide. NotebookLM is the clearest example: upload your files, ask questions, and every answer cites the passage it came from, which you can click to read. Projects in ChatGPT and Claude, and Gems in Gemini, give a similar effect when you attach documents and instruct the model to use only those.

Two rules apply to both:

1. Grounded is not the same as correct. The model can misread a source, cite the right page for the wrong claim, or summarize a low-quality site. Open the source and confirm it says what the answer says.
2. Prefer sources you would have trusted anyway. A grounded answer that cites a content farm is a well-cited bad answer. Ask the tool to prefer primary sources, and check the domain of every link.

A worked example: checking a market-overview paragraph

A marketing manager asks a chat assistant for "a short overview of the US market for commercial coffee equipment" to open a proposal. The draft is smooth: a market size figure attributed to a research firm, an annual growth rate, a named leading vendor, and a quote from an industry executive.

Applying the process:

1. Triage: it is going to a client. Full check.
2. Extract: four claims: market size with source, growth rate, leading vendor, executive quote.
3. Verify by type: the market size figure cannot be found on the research firm's site or anywhere else with that number; the growth rate turns up in a press release for a different, broader category; the vendor is real and plausibly leading, which a trade publication confirms; the quote appears nowhere.
4. Uncertainty pass: rerunning the draft through the editor prompt above labels three of the four claims UNSUPPORTED, which would have saved the searching if it had been done first.
5. Grounded tool: a search-grounded query returns a dated industry report summary with a different market size figure and a link to the report page.

Result: the paragraph is rewritten around the one sourced figure, the vendor is kept with a citation, the growth rate is dropped, and the quote is replaced with a linked statement from the report. Fifteen minutes, and the proposal now contains nothing the manager cannot defend.

Privacy while you verify

Fact-checking often means pasting drafts, sources, and data into tools. The same rule applies as everywhere else: never paste confidential data into a consumer AI tool unless your organization has approved it. That includes client names, financials, contracts, and anything covered by HIPAA, attorney-client privilege, PII rules, or an NDA. Strip identifiers before pasting, or use the enterprise tool your employer licenses. AI privacy at work has the full set of rules.

The fact-check checklist

Save this in your prompt library and run it before anything AI-assisted leaves your hands.

• Triage: I know where this output is going and checked at the level that destination requires.
• Claims extracted: every factual statement is on a list, not buried in prose.
• Facts and dates: each traced to a primary source with a date on it.
• Numbers: each has a named source; every calculation recomputed.
• Quotes: each found verbatim in the original.
• Citations and links: each opened; title, author, and year match; the source actually supports the claim.
• Code and formulas: run and tested on known cases; all functions and libraries exist.
• Summaries: at least three specifics spot-checked against the source document.
• Uncertainty pass: the model labeled its claims; every UNSUPPORTED item was checked or cut.
• Grounding: anything current came from a search-grounded tool with a dated source I opened.
• Privacy: nothing confidential went into an unapproved tool.
• Signature test: I would defend every sentence to a skeptical reader.

Next steps

• Understand why the errors happen in AI hallucinations explained and How large language models work.
• Protect data while you check with AI privacy at work, and write the verification step into your team's rules with How to write a team AI policy or the AI policy generator.
• Pick grounded tools: Perplexity for research and NotebookLM for your own documents.
• Summaries are the most common thing to check; see How to summarize documents with AI.
• Glossary: hallucination, grounding, retrieval-augmented generation, and human in the loop.
• The course turns this into a repeatable habit with exercises for each step.