How to Fact-Check AI Output Before Publication

An AI-generated citation is a lead, not evidence. Models can invent sources, attach a real link to the wrong claim, merge facts from different dates, remove a crucial qualifier, or calculate from incompatible numbers. Publication requires a claim-level verification process.

Create a Claim Inventory

Mark every statement that is specific, current, quantitative, comparative, consequential, surprising, or attributed to another person. Also mark recommendations whose safety depends on a factual premise.

Prioritize by potential harm:

1. medical, legal, financial, safety, security, and rights-related claims;
2. accusations, quotations, and claims about identifiable people;
3. prices, plans, product capabilities, laws, and current events;
4. statistics, benchmarks, and comparative statements;
5. stable background facts.

High-risk material may require a qualified subject-matter reviewer, not merely another web search.

Find the Primary Source

Use official documentation, legislation, court or regulator records, original research, datasets, filings, standards, and direct statements where possible. A search-result snippet, AI overview, or article repeating another article is not the original evidence.

Record source title, publisher, URL, publication date, access date, and exact supporting location. For documents, include page or section. For rapidly changing pages, preserve an authorized archive or screenshot according to editorial policy.

Verify Entailment, Not Keyword Match

Ask whether the source supports the exact wording. If documentation says a feature is available to selected users in preview, the article cannot say it is generally available. If research finds an association in one population, the article cannot claim universal causation.

Preserve qualifiers: may, observed, estimated, preliminary, under specific conditions, or as of a certain date. Stronger prose is not more accurate prose.

Check Time, Scope, and Identity

Confirm that evidence refers to the same product version, plan, region, population, and period. Organizations and products can share names. A model capability in an API may not exist in a consumer application. A global policy may have jurisdiction-specific exceptions.

Add an “as of” date to volatile claims and define a refresh trigger before publication or purchase.

Audit Numbers

Trace each number to a source table or calculation. Preserve units, currency, base year, sample size, denominator, and whether values are nominal, real, annualized, cumulative, forecast, or observed.

Recalculate percentages and totals independently. Watch for percentage versus percentage-point changes, averages hiding unequal groups, and charts with truncated axes. If two sources use different definitions, do not combine them without a documented normalization.

Confirm Quotations

Locate the original transcript, recording, publication, or official statement. Verify the exact words and sufficient surrounding context. Check speaker, date, translation, and whether edits change meaning. If the original cannot be found, paraphrase cautiously with attribution or remove the quotation.

Test Product and Technical Instructions

Open documentation for the stated version and run commands or steps in a safe environment. Verify prerequisites, permissions, output, and failure cases. Never publish generated code that has not been reviewed for security, licensing, and compatibility.

For screenshots, confirm that the interface is current and does not expose private data. For generated images, verify that captions do not imply a real event or person when the visual is synthetic.

Use a Verification Table

Claim	Risk	Primary source	Support	Action
Feature is generally available	Medium	Official release note	Says limited preview	Revise
Study reduced errors by 18%	High	Original paper, Table 2	Correct for one cohort	Add scope

This table makes editorial review reproducible and prevents corrected claims from returning in a later draft.

Run an Adversarial Pass

Ask an AI system to identify sentences likely to be stale, overbroad, unsupported, or easily misread. Use the result as a checklist, not as verification. A second model can repeat the first model's error.

Have a human reviewer examine the central argument: even individually correct facts can be selected or ordered to support a misleading conclusion.

Publish With Transparent Uncertainty

State what is known, what is inferred, and what remains unknown. Link directly to material sources. Correct mistakes visibly according to the publication's correction policy.

Fact-checking is not a final spell-check. It is a traceable process that connects every important public claim to evidence appropriate to its risk, scope, and date.