Separate fluency from evidence
An AI answer can be easy to read, logically arranged, and wrong. It may combine accurate background with an invented statistic, misquote a real paper, or cite a page that does not support its conclusion. Verification means checking the claims against evidence, not judging how confident the wording sounds.
OpenAI warns that ChatGPT can produce incorrect facts, fabricated references, and overconfident answers. Search tools can help find evidence, but the presence of a citation is not itself a guarantee of correctness. 1
Use the workflow below for research notes, business drafts, and everyday decisions. For medical, legal, financial, or safety-critical questions, consult an appropriately qualified professional and authoritative sources rather than treating this checklist as a substitute.
Triage by consequence
Not every statement requires the same effort. A suggested name for a fictional café is low risk. A claim about tax deadlines or medication interactions deserves a much higher standard. First ask what happens if the answer is wrong and who could be affected.
| Claim type | Initial verification route | Common failure |
|---|---|---|
| Product feature | Current official documentation | Old plan details presented as current |
| Statistic | Original report and methodology | Missing denominator or wrong population |
| Research finding | Paper, methods, and limitations | Correlation described as causation |
| Quotation | Original text or recording | Invented or altered wording |
| Calculation | Independent recomputation | Wrong units or hidden assumptions |
| Legal requirement | Official jurisdiction-specific source | Advice from the wrong location |
Prioritize claims that determine your decision. If the recommendation depends on a product being free for commercial use, verify that condition before polishing the rest of the document.
Build a claim ledger
Copy the important claims into a small table. Give each one a proposed source, a status, and a note about limitations. Useful statuses are supported, partially supported, contradicted, and not verified.
For example, an assistant might say: “This service gives all users unlimited exports.” Split that into questions: Which service? Which account type? What counts as an export? Is there a fair-use restriction? Does the statement apply now?
Do not group several claims under one citation if the source supports only one of them. A product page describing PDF export does not establish unlimited exports, commercial licensing, or privacy compliance. Verification becomes easier when each sentence makes a precise, bounded claim.
Confirm that the citation exists
Open each cited URL yourself. Check the domain, page title, author or organization, and publication information. Look for a normal official address rather than a convincing imitation or unrelated mirror.
For academic references, search the exact title through your library, the publisher, or a recognized scholarly index. Confirm the authors, journal, year, and identifier. A real DOI paired with an unrelated title is still an invalid citation.
If a link fails, search the organization or title independently. A broken URL does not prove that the underlying claim is false, but it leaves the evidence unresolved. Do not mark it verified merely because a similar-looking page appears in search results.
Check whether the source supports the claim
Read the relevant passage and enough surrounding text to understand its conditions. Search snippets and AI summaries can omit exceptions. A heading saying “free access” may lead to terms restricting particular features, locations, or users.
Suppose an answer says a study proves that a learning app improves exam results. The paper might actually report a survey of user satisfaction. Both concern education, but satisfaction is not an exam-performance measurement. Record what the study measured, not what the assistant inferred.
Distinguish direct evidence from interpretation. You may reasonably draw a conclusion from several sources, but label it as your analysis. Do not attribute your inference to a source that never makes that claim.
Inspect dates, scope, and methodology
For current facts, check when the source was updated and whether it applies to the current product or policy. For historical facts, an older primary document may be exactly what you need. Newer is not always better; relevance to the claim is the deciding factor.
For a statistic, identify the population, sample size when available, collection period, and definition of the measured outcome. “Half of respondents” is not the same as half of all people. A percentage without its denominator is difficult to interpret responsibly.
Check jurisdiction and version. A rule in one country may not apply in another, and documentation for an older software release may describe behavior your installation no longer has. Record these boundaries alongside the evidence.
Recalculate quantitative answers
Write down the inputs, units, and formula. Recompute with a calculator, spreadsheet, or a short independently reviewed script. If an assistant says a price rose from 80 to 100 by 25 percent, calculate the difference, 20, divided by the original 80.
Watch for percentage points versus percentage changes. An increase from 10 percent to 15 percent is five percentage points and a 50 percent relative increase. Both are valid descriptions, but they communicate different things.
For tables, check whether totals reconcile and whether rounding explains small differences. Never accept a calculated result solely because it is displayed with many decimal places. Precision in formatting does not establish accuracy.
Handle disagreement without forcing certainty
If two sources disagree, compare their dates, definitions, populations, and authority. They may be answering different questions. Prefer original evidence for the specific claim rather than counting how many secondary sites repeat a statement.
When uncertainty remains, narrow the wording or omit the claim. “The provider lists this feature for eligible accounts” is more defensible than “Everyone gets it.” For a consequential unresolved issue, pause the decision and seek expert clarification.
Do not use another AI answer as the only independent check. A second system may repeat the same mistaken material. It can suggest search terms or possible weaknesses, but the evidence still needs inspection.
Publish with a review trail
Keep the source URL, relevant passage, access date when useful, and your final wording. This makes later updates easier and lets another reviewer understand the decision. Avoid attaching a citation to a paragraph whose stronger claims it does not support.
If you are choosing research tools, explore ChatGPT Plus alternatives and the free tools category, but evaluate their answers with the same standards. Paid access is not a substitute for verification.
Finish with a final sweep: every consequential claim has evidence, every quotation matches, every calculation is checked, and every unresolved uncertainty is visible. Responsible use is not about proving that AI never makes mistakes. It is about preventing an unchecked mistake from becoming your published fact.