How to Fact-Check AI Answers (A 6-Step Method)

AI tools write fluent, confident answers. They also get names, dates, numbers and sources wrong, and the wrong ones read exactly like the right ones. The fix is a repeatable check. This is the 6-step method: split the answer into claims, rank them by risk, check each against a primary source, open every citation, record what you verified, and keep a final human sign-off.

Why AI answers need checking

An AI model generates text by predicting what a good answer looks like. It does not look facts up unless a tool you are using is built to search and cite. That is why it produces a fake statistic or a book title that does not exist with the same tone as a true answer. This is often called a hallucination.

The risk is highest when the answer contains specifics: numbers, names, dates, quotes, legal or medical statements, prices and links. General advice is lower risk. Anything you plan to publish, send to a client or base a decision on needs a check.

The 6-step method

Step 1: Split the answer into individual claims

Read the answer and underline every statement that could be true or false. One sentence often holds three claims. "The company was founded in 2011 and now has 4,000 employees" is two claims.

Write them as a list. You cannot check a paragraph. You can check a claim.

Step 2: Rank claims by risk

Mark each claim high, medium or low.

• High: numbers, statistics, quotes, names, dates, legal, medical or money statements, anything you will publish.
• Medium: how a tool or process works, comparisons, rankings.
• Low: common-sense advice, definitions you already know.

Start with the high-risk claims. If time is short, check those and cut or soften the rest.

Step 3: Check each claim against a primary source

A primary source is where the fact originates: the company's own filing or website, a government dataset, the original study, the official documentation. Search for the claim yourself and open the original, not a blog that repeats it.

Use two sources when the claim matters. If you cannot find the claim in any source, treat it as unverified and remove it.

Step 4: Open every citation and find the quote

When an AI gives a link, a paper title or a case name, verify it in three moves. Open the link. Confirm the page exists and is what the AI said it is. Find the exact sentence that supports the claim.

If the link is dead, the title returns nothing in a search, or the page says something different, the citation is fabricated or wrong. Delete the claim or replace it with a real source you found.

Step 5: Ask the AI to show its work, then check that too

Ask: "List each factual claim in your answer and the source for each." Ask: "Which of these claims are you least sure about?" These prompts expose weak spots quickly.

Treat the response as a to-do list, not as proof. The AI can invent a source for a claim just as easily as it invented the claim. Every source it names still goes through Step 4.

Step 6: Record what you verified and sign off

Keep a short log: the claim, the source you used, the date you checked it. Two minutes per claim is a realistic budget for high-risk items. The log protects you when someone asks where a number came from, and it makes the next check faster.

A person makes the final call. Read the finished piece once more for claims that slipped in during editing.

A worked example

This example is invented to show the method. It is not a real company or quote.

You ask an AI to summarize a competitor. It replies: "Northwind Analytics was founded in 2011 by two engineers, has about 4,000 employees, and was named a leader in a 2023 industry report."

1. Split: founded in 2011; founded by two engineers; about 4,000 employees; named a leader in a 2023 report.
2. Rank: employee count and the report are high risk if you publish them. The founding year is medium.
3. Check: open the company's About page and a recent filing or press page for founding date and headcount. Search for the report title.
4. Citations: the AI named a report. Search for it. If no such report exists, remove the claim.
5. Ask the AI: "Which claims are you least sure about?" Check the ones it flags first.
6. Record: log each verified claim with its source and the date.

The result is a short, accurate paragraph you can stand behind.

Prompts that make checking easier

• "Separate what you know from what you are inferring."
• "Give me the source for each statistic. If you have no source, write 'no source'."
• "Flag anything in your answer that depends on a date after your training data."
• "Quote the exact sentence from the document I pasted that supports each claim."

These reduce the amount you have to check. They do not replace checking.

When to skip the deep check

You do not need a source for every line. A brainstorm of subject lines, a rewrite of your own paragraph or an outline of ideas needs a read-through, not a research session. Match the effort to the cost of being wrong.

Frequently asked questions