Separate facts, interpretations, and opinions in AI-assisted writing
For each important sentence in an AI-assisted draft, ask what kind of statement it makes, what evidence supports it, and whether the reader can tell the difference. A fact is checkable against evidence; an interpretation explains what evidence may mean; an opinion expresses a judgment or preference. Labeling these roles during editing helps catch unsupported claims, hidden assumptions, and opinions presented as settled truth.
What distinguishes a fact, an interpretation, and an opinion?
**Fact:** A statement that can be checked against records, direct observation, or other evidence. “The report was published on 12 March” is a factual claim if the report itself confirms the date. Facts still need verification: a precise number or confident tone does not make a statement true.
**Interpretation:** An explanation or inference drawn from facts. “The timing suggests the announcement was meant to draw attention to the workshop” interprets a sequence of events; the timing alone cannot establish intent. Interpretations should identify their basis and use language that reflects uncertainty, such as “may suggest,” when the evidence does not settle the point.
**Opinion:** A value judgment, recommendation, or preference. “This layout is confusing” is an opinion unless the article specifies a measurable standard and provides evidence for an evaluative conclusion. Opinions can be useful, but readers should be able to recognize whose judgment they are reading and what reasoning informs it.
These categories can overlap in a sentence. “This layout is wasteful because it uses 20% more paper” combines a judgment (“wasteful”) with a factual claim (“uses 20% more paper”) and a causal or predictive implication. Check and attribute each part separately.
How to review a draft sentence by sentence
Use a simple three-column audit: **claim**, **type**, **support or change needed**. Mark each sentence F (fact), I (interpretation), or O (opinion); split sentences that contain more than one type. Then work through these checks:
A worked example: turn a smooth claim into a supported one
Suppose an AI draft says: “The new schedule improved productivity and made employees happier.” This sentence sounds like a fact, but it bundles several claims: that productivity changed, that the schedule caused the change, and that employee satisfaction rose. A careful edit asks what data exists, who was included, how productivity and satisfaction were measured, and whether other changes could explain the results.
If a dated company report shows output rose during the period, a narrower factual sentence could be: “The company reported higher output during the first quarter after the schedule changed.” That wording attributes the claim and preserves the time period, but it does not establish that the schedule caused the increase. If employee survey results also exist, describe the survey and respondents rather than converting them into a claim that all employees were happier.
An interpretation might then say: “The timing is consistent with the schedule contributing to higher output, although the figures alone do not establish cause.” An opinion or recommendation could follow: “For teams weighing a similar change, the approach may be worth testing if output is the main priority.” The recommendation now has a stated condition; it is not disguised as a universal result.
What to do when the source is weak or the claim is uncertain
A citation is not proof by itself. Open the linked source, check the relevant passage, and note whether it is a firsthand record, a direct statement, or a report repeating another source. Reuters recommends named sources where possible and cautions against misleading sourcing; that guidance is especially relevant when an AI draft produces citations that look plausible but do not substantiate the sentence.
If you cannot retrieve or verify a source, do not leave its claim in declarative form. You can remove the claim, describe it explicitly as unverified only if that uncertainty matters to the reader, or continue research until a suitable source is available. Do not use an AI-generated reference as evidence merely because the model supplied it. Reuters specifically warns journalists not to rely on unverified facts generated by AI and says AI-generated claims require independent checking.
When evidence conflicts, describe the disagreement and identify what each source actually establishes. Do not flatten a disputed interpretation into a settled fact. If the disagreement cannot be resolved within the article’s scope, state the limit plainly or omit the claim.
A compact editing checklist
Before an AI-assisted article is ready for editorial review, ask:
The aim is not to label every sentence for readers. It is to make the distinctions during editing, then write clear prose that lets readers see what is established, what is inferred, and what is being argued. The author or editor remains responsible for checking the article’s claims and sources before publication.