Fact-check AI-generated content one claim at a time

A real source can sit beside a sentence it does not support. The link works, the report discusses the right subject, and a few words in the draft still go beyond what anyone established.

To fact-check AI-generated content, take each material claim back to the passage that is supposed to support it. Compare the exact wording, scope, and date before deciding whether the sentence can stay.

Follow the claim into the source

Consider this hypothetical draft sentence:

AI summaries cut support response times.

The draft cites a pilot report. For this example, the report and its findings are also invented. Its relevant passage says:

During a two-week pilot, one support team's median time to first reply was lower than in the previous two weeks. The team introduced AI summaries and a new triage queue together. The pilot did not separate their effects.

The report supports an observation about one team over a particular period. The draft makes a causal claim about AI summaries.

Open the report and read the surrounding explanation. In this example, the change to the triage queue matters: the report does not tell us which intervention affected reply times. Nor does one team's pilot establish what happens across support teams generally. Even the metric needs care. “Time to first reply” describes something more specific than an undefined “response time.”

A supported rewrite would be:

In one support team's two-week pilot, median time to first reply was lower than in the previous two weeks. The team introduced AI summaries and a new triage queue together; the pilot did not establish their separate effects.

That wording keeps the observation and the information needed to interpret it. If the article needs to establish that AI summaries caused the improvement, this source cannot finish the job.

Adding “may have” to the original sentence would leave the causal explanation unresolved. Decide whether the article needs that explanation, then find evidence suited to it or remove it. Do not use a softer verb to hide the missing check.

Ask five questions of each material claim

A sentence can contain several propositions. Words such as stores, exposes, prevents, and guarantees each assert behavior that may need different evidence. Split the sentence when that makes the review clearer.

For each claim, ask:

1. What does the sentence literally say? Include the implications of words such as “all,” “always,” “only,” and “because.”
2. Does the source cover that system, population, version, and period? A finding about one pilot or an older release has a narrower reach than a general present-tense claim.
3. Is the source reporting a fact, an interpretation, or a recommendation? Advice to adopt a practice does not establish that someone adopted it or that it worked.
4. Does the wording claim causation where the evidence shows only association or sequence? In the example, faster replies followed two simultaneous changes.
5. Which uncertainty has to remain attached to the sentence? Keep the limits a reader needs to understand the finding, especially when removing them would change the conclusion.

Numbers, dates, definitions, product capabilities, policy requirements, safety claims, and causal explanations deserve this attention. A transition usually needs less. Match the depth of review to the consequence of being wrong.

Keep a small record beside the draft: the exact claim, source and relevant passage, source date or version, what supports or contradicts the wording, and the editorial decision. A URL alone leaves the next reviewer to find the passage and infer why you accepted it.

NIST's Generative AI Profile recommends documenting fact-checking and reviewing sources and citations as part of managing generative AI risks. That supports keeping an inspectable record; it does not establish that a particular checklist eliminates errors.

Give the claim a decision

After comparing the sentence with its evidence, choose what happens to it:

• Keep it. The source supports the wording, scope, and date.
• Qualify it. A narrower statement is supported. In the hypothetical pilot, keep the observed change and the limits on attributing its cause.
• Remove it. Support is absent or contradictory, and the article does not need to pursue the claim further.
• Escalate it. The claim matters, but a stronger source or someone with relevant expertise must resolve it.

Write the repair where possible. “Overstated” tells the author less than identifying the unsupported cause and supplying the narrower observation.

For medical, legal, financial, security, and other consequential subjects, the decision may need a qualified domain reviewer. Preserve the question they must resolve. A general editor can identify missing support without being qualified to decide the underlying issue.

Check provenance and disclosure for their own purposes

Factual review asks whether evidence supports a claim. Provenance asks what is known about the content's origin and history. Disclosure asks what readers should be told about its author, production, or purpose.

C2PA's explainer describes Content Credentials as evidence about origin, history, and integrity. That information alone cannot establish that the content is factually accurate. Missing credentials do not, by themselves, establish that it is false either.

Google's people-first content guidance encourages clear authorship and useful explanations of how content was produced, including automation disclosures when readers would reasonably expect them.

Those checks can matter for the same article. Neither an authorship label nor a detector score settles whether the pilot report supports the draft's causal claim. The sentence still needs its source comparison.

Let another reviewer inspect the decision

Give the reviewer the original sentence, proposed wording, exact source location, and reason for the change. Ask them to check the source itself and identify any part of the revised claim it still cannot support.

In the pilot example, they should be able to trace the narrower sentence to the observation and see why the causal claim was removed. Agreement between two reviewers is not proof of truth. If they disagree, record the disputed clause and the evidence or expertise needed to resolve it. The author then has a specific question to answer before that claim is ready.