How to verify a chatbot claim before you use it

A chatbot answer can be well written, confident and still combine one accurate fact, an outdated date and a source that does not exist. You do not need to learn to spot AI writing to use it carefully. You need to learn to check the claim you intend to reuse. In ten minutes, you can apply this method to a figure for a report, a quotation for a presentation, a recommendation for a client or a rule you are about to share with other people.

The central idea is simple: do not ask the chatbot whether its answer is correct. Treat it as a draft handed to you by someone without a signature. OpenAI advises users to directly check quotations, data, technical information and references when accuracy matters. Your job is not to prove that a model “lies”; it is to decide whether this particular claim meets the evidential standard required for your use.

The 60-second method

Before copying anything, write down the exact part you plan to use. Not “the answer about the grant,” but “the application closes on 30 September,” “the report attributes a 12% saving,” or “the rule requires this document.” If a sentence contains three facts, split them up: each can have a different source and a different period of validity.

Then classify the cost of being wrong. For a headline idea or an informal explanation, a quick check may be enough. For money, health, employment, tax, law, safety, or a decision that affects another person, the chatbot is not the final source: you need the document, register, study or authority that supports the statement. NIST’s Generative AI Profile defines confabulation as erroneous or false content presented confidently. That is why a confident tone is not evidence.

With the claim written down and the risk clear, ask four questions: who could know this first-hand? Where does that person or organisation publish? When was it published or updated? Does it say exactly what I want to claim? If one answer fails, do not use the sentence yet.

Step 1: turn the answer into a checkable record

Open a note and make four fields: claim, date, primary source, and verdict. Copy the claim without embellishment. If the chatbot wrote, “the company drastically cut emissions thanks to AI,” ask for, or write, a measurable version: “the company says it cut emissions by X% between A and B through C.” Only then do you know what to look for.

The date is not a stylistic detail. A 2024 number may be accurate but useless for a 2026 decision; a rule may have been announced but not yet apply; a software feature may have changed plan or name. Record the date provided by the chatbot and do not accept it until you see it in the original document. Where there is no date, mark the record “validity not confirmed.” This stops an apparently timeless statement from quietly relying on an old version.

In the primary-source field, write the document you expect to find: an authority’s page for an obligation, the research paper for a scientific result, a company report or announcement for its own product, an official register for a corporate fact. A news report can help you find it, but it does not automatically replace the document it summarises.

Step 2: search for the document, not the chatbot’s sentence

Do not paste the whole answer in quotation marks into a search engine. That often leads to pages repeating it, aggregators, or other generated text. Search the proper names in your record: the agency, programme, report title, authors, version, year, or reference number. Add the document type you need: “PDF,” “report,” “decision,” “paper,” or “documentation.”

If the chatbot gave you a link, open it before trusting it. Check three things: the domain belongs to the supposed author, the page exists, and its title or text supports the sentence. An official-looking URL may be a press page summarising another source; an academic-looking citation can point to a different article; a link can be dead. An AI provider does not make a reference genuine simply by rendering it convincingly.

Here is a useful rule: the larger the conclusion, the closer you need to get to the origin. For “a ministry announced X,” find the ministry announcement. For “a study found X,” open the study and locate the table, method or conclusion. For “a product can do X,” use its documentation or terms, not a third-party comparison. NIST’s AI Risk Management Framework recommends managing risks in their context of use; in practice, the same answer deserves different checks when it merely starts a search and when it triggers an action.

Step 3: read for the boundary, not just the match

Finding a similar sentence does not complete the check. Read the preceding and following paragraph, the date, definitions of terms, and every condition. AI answers often drop precisely the limit that changes the meaning: “in a particular sample,” “up to,” “for eligible customers,” “according to the company’s estimates,” “in a pilot,” or “not a substitute for professional assessment.”

Try this: underline the words that narrow the claim in your record. If the original says “may,” do not write “does.” If the report says “the authors estimate,” do not write “it has been proven.” If a company reports its own result, attribute it: “according to the company.” That is not defensive filler; it tells the reader where the figure comes from and what kind of check it has passed.

Also verify the unit. “20%” could be a relative improvement, an absolute reduction, a result on a test set, or a forecast. Write beside the number what it measures, what it is compared with, and over which period. If you cannot fill in those three boxes, the number is not ready to circulate.

Step 4: add a second check when the impact calls for it

A primary source answers, “what does the author or authority say?” It does not always answer, “how robust is it?” or “how does it apply to my case?” When you are going to base an important recommendation on the claim, add an independent and competent second source: a publication analysing the study, a supervisory authority, a professional association, or the applicable legal text.

This is not an exercise in collecting links. Find a source that can contradict or limit the first one. A company may describe a tool; a contract describes its conditions; an authority describes obligations; a research paper describes its results; a review describes its limits. If the sources agree, note what they agree on. If they conflict, do not choose the one that sounds most decisive: keep the disagreement and narrow your conclusion.

In health, legal, financial or safety matters, this step does not turn a chatbot into a professional adviser. It helps reveal that you need the relevant professional, authority or documentation before acting. AI is useful here for framing better questions and locating documents; it does not replace the responsibility of the person making the decision.

Step 5: decide with three labels

At the end of the record, do not simply write “true” or “false.” Use one of three labels. Usable: the primary source, date and scope support the exact sentence. Usable with attribution or a caveat: the document supports something close, but you must name who says it, add the date, or keep a condition. Do not use: there is no source, the source does not say it, validity is uncertain, or the risk requires a check you have not been able to make.

This output matters because it avoids two opposite mistakes: passing on a doubtful sentence because you fear losing time, or deleting a useful idea because it is not absolute certainty. Most good chatbot use does not end with “100% confirmed.” It ends with a more precise sentence: “according to the report from…,” “in the version published on…,” “for this sample…,” or “I could not confirm this point.”

Do it now with a real answer

Choose one chatbot sentence you were about to copy today. Set a ten-minute timer and fill in the record. Do not start with a medical, legal or financial topic: try a software feature, an application deadline or a historical fact. Then rewrite the sentence at the level of precision the source permits.

If you cannot find the original document in those ten minutes, do not force a result or ask the model for a replacement citation. Leave it out, mark it “do not use,” and move on to another part of the work. That is the durable habit: a persuasive answer does not become knowledge through repetition. It becomes usable information when you can show which document supports it, what date it has, and where its limits lie.