An AI answer can be clear, specific, and completely wrong. Confidence is a writing style, not evidence. The NIST Generative AI Profile calls confidently presented false content “confabulation” and notes that invented logic and citations can make a bad answer look justified.
Fact-checking does not mean asking the chatbot, “Are you sure?” That is the same witness reviewing its own testimony. It means taking the answer apart and checking its important claims against evidence outside the conversation.
Use the TRACE method below. It is designed for ordinary AI answers, not just academic papers: travel details, product comparisons, workplace research, health explanations, legal summaries, historical facts, and statistics.
T — Triage the Risk
Decide how much checking the answer deserves before you start. Verification should match the cost of being wrong.
| Risk | Examples | Minimum response |
|---|---|---|
| High | Medical care, legal rights, taxes, investments, physical safety, accusations about a person | Check primary authorities and current dates; consult a qualified professional before acting |
| Medium | Work submitted to a client, purchases, travel plans, public posts, school assignments | Verify every claim that could change the decision or damage credibility |
| Low | Brainstorming, tone changes, fictional writing, rough organization | Check only factual details you keep or publish |
Raise the risk level when an answer says latest, always, proven, illegal, safe, guaranteed, or everyone. Also raise it when the answer asks you to act quickly. A restaurant recommendation being outdated is annoying; an outdated visa rule can end a trip.
For high-risk questions, a fact-checked chatbot answer is still not a diagnosis, legal opinion, or financial recommendation. Verification helps you ask better questions; it does not create professional accountability.
R — Reduce the Answer to Checkable Claims
Do not try to verify a six-paragraph answer as one block. Turn it into a claim ledger. Mark:
- names, dates, numbers, quotations, and links;
- claims that a rule, study, product, or event exists;
- causal claims: “X causes Y”;
- comparisons and superlatives: “cheapest,” “first,” or “most effective”;
- recommendations that depend on hidden facts;
- statements that may have changed since publication.
Split compound sentences. “The policy began in 2023 and applies to every contractor” contains at least two claims. One can be true while the other is false.
You can use AI for the clerical step without letting it judge itself:
Copy the answer below into a claim ledger.
For each factual claim, list:
1. the exact claim;
2. claim type (date, number, quote, policy, causal, other);
3. consequence if wrong (high, medium, low);
4. whether it is time-sensitive.
Do not verify, correct, combine, or add claims.
[paste answer]
Start with the claims that are both consequential and easy to disprove. If the answer depends on a nonexistent court case or a study with the wrong title, you have learned something important about the whole response.
A — Anchor Each Claim to the Right Source
A source is useful only when it fits the claim. “A university website” is not automatically an authority on every subject, and a company page can establish what the company promises without proving that the promise works.
| Claim | Best starting point |
|---|---|
| Law, deadline, benefit, or government policy | Statute, court opinion, agency page, or official notice for the correct jurisdiction |
| Scientific or medical finding | Original study plus a recent systematic review or official clinical guidance |
| Statistic | Original dataset and its methodology or definitions |
| Quote | Full transcript, recording, filing, or original publication |
| Product feature or price | Current official documentation, then an independent test if performance matters |
| Company claim | Company filing or announcement for what it says; independent evidence for whether it is true |
| Breaking news | Primary announcement or public record plus reputable reporting that adds context |
Open every citation the chatbot supplies. Confirm that the page exists, the title and author match, and the source actually supports the sentence. Use Find in Page for a distinctive phrase, number, or name. A real article that discusses the topic but does not establish the claim is not supporting evidence.
Treat an AI-generated citation as a search lead, never as proof. For papers, verify the title, authors, year, journal, and DOI in the publisher record or a trusted scholarly database. For web pages, inspect the domain carefully; a familiar-looking logo does not make a look-alike URL official.
C — Cross-Check Outside the Answer
Open a fresh search rather than following only the chatbot's suggested trail. Search the exact claim, then vary the wording. Useful patterns include:
"distinctive phrase from the claim"
topic keyword site:gov
report title filetype:pdf
claim keyword correction
claim keyword criticism
The site: and filetype: operators narrow discovery; they do not certify accuracy. Google also warns that reliable material can be scarce for a new or fast-changing topic, so a thin results page is a reason to wait or widen the search—not permission to fill the gap with certainty. See Google Search Help on reliable results.
Check the source laterally: leave the page and see what credible outsiders say about its publisher, expertise, funding, and record. Google's “About this result” guidance can surface basic source context, while Cornell University Library's media-literacy guidance recommends cross-checking with multiple reputable sources.
Two sources are not necessarily two confirmations. Ten pages may copy the same press release, wire story, or incorrect post. Look for independent evidence, not repeated wording. Conversely, one primary source may be decisive: the current regulation itself is stronger evidence of its text than five blog summaries.
Try to disprove the answer. Search for an exception, correction, retraction, newer version, or contrary result. Then check scope: population, location, date range, study design, and definition. Many AI errors are not fabricated from nothing; they are real facts moved into the wrong context.
E — Explain the Verdict and Keep the Evidence
Avoid forcing every claim into “true” or “false.” Use verdicts that preserve what you learned:
- Supported: the source directly establishes the claim in the stated scope.
- Supported with limits: the core is right, but important conditions were omitted.
- Outdated: it was once supported but newer evidence or policy supersedes it.
- Unsupported: you found no adequate evidence after a reasonable search.
- Contradicted: stronger evidence says the claim is wrong.
- Unverifiable: the needed record is private, missing, or not yet available.
Record the claim, verdict, source link, source date, date checked, and one sentence about scope. This small evidence log prevents you from repeating the search later and lets another person audit your conclusion. “Unsupported” should describe your search result, not pretend to prove that something never happened.
Worked Example: “AI Help Means No Copyright”
Suppose a chatbot says:
“No. If you used a chatbot to edit your short story, the entire story is AI-generated and cannot be copyrighted in the United States.”
Triage: This affects legal rights, so treat it as high risk and use the responsible U.S. authority.
Reduce: The response contains three claims: any AI use makes the whole work AI-generated; AI-assisted work cannot receive copyright protection; and the rule applies categorically in the United States.
Anchor: The U.S. Copyright Office's January 2025 copyrightability report announcement says that using AI to assist creation—or including AI-generated material in a larger human-created work—does not by itself bar copyrightability. It says protection depends on sufficient human-authored expressive elements; merely providing prompts is not enough to claim authorship of machine-determined output.
Cross-check: The Office's registration guidance adds the missing distinction: human-authored selection, arrangement, or modification may be protected, while material whose expressive elements were generated by AI is not. Applicants also need to disclose appreciable AI-generated content when registering a work.
Explain: Verdict: contradicted and overbroad. A careful replacement is: “Using AI as an editing tool does not automatically remove copyright protection from your human-authored story. Protection and registration depend on which expressive elements a human created, and appreciable AI-generated material may need to be disclosed and excluded from the claim.” That is still general information, not advice about a particular registration.
Notice what fixed the answer: not a better-sounding prompt, but claim separation, the correct jurisdiction, primary sources, and attention to exceptions.
Common Shortcuts That Fail
“The chatbot gave five links.” Quantity does not establish relevance. Open them and trace each claim to the text.
“A second chatbot agreed.” Models may share training material, search results, or the same popular misconception. Agreement is not independent evidence.
“It came from a .org or a polished PDF.” Domain endings and design are weak signals. Identify the author, publisher, method, date, and incentive.
“The arithmetic looks reasonable.” Recalculate totals, units, percentages, and denominators. Check whether “percent” was confused with “percentage points.”
“The screenshot proves it.” Images can be old, cropped, edited, or attached to the wrong event. Google's About this image can show earlier appearances and other pages using a similar image, but Google cautions that metadata can also be modified.
The 90-Second Publication Check
Before you send, submit, buy, book, or post based on an AI answer, ask:
- [ ] Did I isolate the claims that matter?
- [ ] Did I open the original sources rather than trust the citations?
- [ ] Do the sources say what the answer says?
- [ ] Are the date, jurisdiction, population, and definitions the same?
- [ ] Did I look for an exception, correction, or newer version?
- [ ] Are my confirmations genuinely independent?
- [ ] Did I label uncertainty instead of hiding it?
- [ ] For a high-risk decision, did I consult the appropriate professional or authority?
For a deeper evidence workflow, continue with How to Use AI for Research. If the problem begins before the answer—because the prompt is vague—see How to Write Better AI Prompts. The habit to keep is simple: use AI for speed, and use evidence for trust.