How to fact-check AI answers: a six-step method you can use every time

An AI assistant can write a polished, well-organised answer in a confident tone — and still include a false fact, a made-up quote or a link that leads nowhere. That isn't a rare glitch. Language models produce text by predicting which words fit, not by looking facts up in a database. So checking what you use is your job.

The good news: you don't need to be an expert. You need a repeatable routine that separates what you can accept as-is from what needs outside confirmation.

#The short answer

To fact-check an AI answer: 1) pick out the claims that can actually be checked, 2) ask the AI to separate what it knows from what it's guessing, 3) confirm each important fact against a primary source, 4) actually open the links and citations, 5) redo the maths and test the code, and 6) scale your effort to what's at stake.

#Why AI gets things wrong while sounding right

The most common chatbot errors fall into a few patterns:

• Hallucinations: names, dates, numbers or references the model generates because they "fit" the text, even though they don't exist.
• Stale information: every model has a training cut-off. Without live web search, it may miss recent changes to prices, laws or software versions.
• Blended sources: it merges two similar things — two products, two countries' rules — into one answer.
• Agreeableness: if your question contains a false assumption, the model often goes along with it instead of correcting you.

OpenAI's own help centre warns that ChatGPT can produce incorrect or misleading information and advises users to verify it. The same caution applies to every assistant, not just one.

#Step 1: Separate checkable facts from opinion

Before checking anything, reread the answer and sort its content into three buckets:

1. Checkable facts: dates, figures, names, technical requirements, direct quotes, laws, specifications.
2. Reasoning: "therefore", "this means that…". You don't need a source here — you need to check the logic holds.
3. Opinions and recommendations: "the best option is…". These aren't true or false; judge whether the criteria make sense for you.

Only the first bucket needs outside sources, and it's usually smaller than it looks. Focusing on it saves a lot of time.

#Step 2: Ask the AI to audit itself

Before you go searching, ask a follow-up in the same conversation:

Review your previous answer. List every factual claim you made and, for each one, say whether it's well established, something that may have changed since your training cut-off, or something you're unsure about.

This doesn't replace verification — the model can misjudge itself too. But it usually flags the weak spots and gives you a prioritised checklist.

#Step 3: Check against primary sources

For each important fact, go to the source closest to where the information originates:

Professional fact-checkers use a technique called lateral reading: rather than studying one page in depth, open several tabs and see what independent sources say about the same claim. If nobody else confirms it, be suspicious.

#Step 4: Actually open the links and citations

References are one of AI's weakest points. Always check:

• That the link works and doesn't return an error.
• That the page says what the AI claims. A real link may not contain the statement attributed to it.
• That a quote is verbatim. Search the exact phrase in quotation marks; if it appears nowhere, it's likely a paraphrase or an invention.
• That books, papers and authors exist. Search the exact title in a library catalogue or Google Scholar.

#Step 5: Redo the maths and test the code

Language models aren't reliable calculators, especially with multi-step arithmetic, percentages or unit conversions.

• Calculations: redo anything important in a calculator or spreadsheet. If the AI gives you a final number, ask for the intermediate steps and check each one.
• Code: run it in a test environment before using it for real. Make sure the functions and libraries it calls exist in the version you have installed — models sometimes invent plausible-sounding function names.
• System commands: don't run commands you don't understand, especially ones that delete or move files or change permissions.

#Step 6: Match the effort to the risk

Not every answer deserves the same scrutiny. A practical rule of thumb:

• Low risk (title ideas, summarising text you pasted in yourself): a quick read is enough.
• Medium risk (a fact for an assignment, setup instructions for an app): confirm the key facts against an official source.
• High risk (health, money, legal matters, security): treat the answer as a starting point and confirm everything with official sources or a qualified professional.

The US National Institute of Standards and Technology (NIST) builds its AI Risk Management Framework on a similar idea: the level of control should be proportional to the impact an error could have.

#Red flags in an AI answer

Look more closely when you see:

• Very precise figures with no source ("73.4% of users…").
• Direct quotes attributed to specific people.
• References to recent versions, prices or regulations.
• Answers that change when you rephrase the question.
• Links with odd structures or to domains you don't recognise.

A simple test: ask the same question twice, in separate chats or with different wording. If the answers contradict each other on specific facts, at least one is wrong.

#Common verification mistakes

• Asking the same AI whether it's right and stopping there. It tends to double down — or to change its answer simply because you pushed back.
• Checking with a second AI. Two models can share the same mistake. Use it to spot disagreements, not to confirm facts.
• Trusting a source just because it exists. Check its date, and whether it's the original source or a copy.

#The bottom line

AI is great for getting oriented, structuring ideas and saving time, but it isn't a source. Treat it as a smart first draft: have it separate what's solid from what's shaky, check important facts against primary sources, and scale your effort to the stakes. With practice the whole routine takes a few minutes and catches most mistakes.

To cut errors at the source, learn to write better prompts and review the most common chatbot mistakes. For longer projects, see how to research a topic with AI without losing track of your sources.