Trace the number back to the original source that actually collected or calculated it, not a summary of it, and check the year, the sample, and any stated caveat before you use it. An AI-generated draft can attach a precise, confident-sounding figure to a sentence with no source behind it at all, or with a real source whose actual number is different, older, or narrower than the sentence implies. The fix is the same whether the number came from an AI tool, a secondhand article, or a conversation you half-remember: find where it originally came from, and read that, not a restatement of it.
Why a specific number is not, by itself, evidence
A statistic phrased precisely, such as "73% of businesses report X," feels more credible than a vague claim, purely because of the precision. That feeling is not evidence of accuracy. An AI tool can generate a specific-sounding percentage with exactly the same process it uses to generate a vague claim, because both are just plausible next text given the context. Treat precision and accuracy as two separate things you have to check separately, not one thing that implies the other.
Step 1: Find where the number actually originated
Search for the specific figure alongside the topic, and look past summary articles to the original source: the survey, the government dataset, the academic study, or the company report that actually produced the number. A blog post that cites "a recent survey" without naming it is not a source. A summary article that reports "researchers found 73%" without linking the actual research is not a source either, it is a claim about a source, one more step removed from the original.
Step 2: Check the year and whether it is still current
A statistic that was accurate in 2019 and repeated without a date in 2026 is being presented as more current than it is. This matters especially for anything related to technology adoption, AI usage, or fast-moving markets, where a few years can make a cited figure meaningfully out of date. Confirm the year attached to the actual data collection, not just the year the article citing it was published, since those are frequently different.
Step 3: Check what was actually measured
A number can be real and still misleading if it is describing something narrower or different than the sentence around it implies. A survey of a specific industry presented as a general finding, or a figure describing intent ("plan to use") presented as behavior ("are using"), changes what the statistic actually supports. Read the original source's own description of its methodology, not just the headline number, to check this.
Step 4: Check the sample and how the data was collected
A statistic based on a small, self-selected survey (people who chose to respond to an online poll, for instance) does not carry the same weight as one based on a large, representative sample or an actual usage log. This does not mean small-sample data is worthless, but it means the confidence with which it gets repeated should match the actual strength of the evidence behind it, not the confidence of the sentence an AI wrote around it.
Step 5: Cross-check against a second source when the stakes are high
For any statistic central to a piece's argument, look for at least one independent source reporting something in the same range. Two credible sources landing in a similar place is meaningfully stronger evidence than one source alone, and a wide disagreement between sources is itself worth reporting honestly rather than silently picking the more convenient number.
What to do when you cannot verify a number
Cut it, or replace it with an honestly framed estimate that says what it is. "Multiple industry surveys report figures in the range of X to Y, though methodologies vary" is honest. Stating a single precise number as settled fact when you could not confirm its origin is not, regardless of how the AI phrased it in the original draft. See how to check whether a study an AI cited actually exists for the citation-existence check this pairs with, and how to fact-check anything an AI writes for you for how this step fits into the full process.
Why this matters more for AI drafts specifically than for human ones
A human writer who half-remembers a statistic usually signals some uncertainty in how they phrase it, a "roughly," an "I think," a note to check it later. An AI-written draft states a half-remembered or fully invented number with the exact same confident phrasing it uses for a number it generated correctly from real training data. There is no tell in the sentence itself. That absence of a tell is the entire reason this needs to be a fixed, repeatable check applied to every number equally, rather than a judgment call about which sentences "feel" like they need it.
A short before-you-publish checklist
- The number traces to an original source you can actually open and read
- The year attached matches when the data was actually collected, not just when an article about it was published
- What was measured matches what the sentence claims was measured
- The sample size and method are known, and the confidence of the sentence matches the strength of the evidence
- A second independent source roughly agrees, for anything central to the piece
FAQ
What if two credible sources report different numbers for the same thing?
Report the disagreement honestly rather than picking whichever number is more convenient for your argument. Naming both figures and their sources, and noting that they diverge, is more useful to a reader than false precision from silently choosing one.
Is a statistic from a company's own report less trustworthy than an independent study?
Not automatically, but it is worth noting the source's own stake in the number when you present it. A vendor-reported adoption figure and an independent academic survey are both usable, provided you attribute each honestly rather than presenting a vendor's own promotional figure as neutral independent research.
How do I verify a statistic when the AI does not cite any source at all?
Treat it as entirely unsourced and search for the claim yourself from scratch, the same way you would if you had encountered the number without any AI involved. An unsourced number in an AI draft is not evidence the number is false, but it carries no more weight than any other unverified claim until you find where it actually comes from.
Does rounding a number change whether it needs verification?
No. Whether a figure is stated as 73% or rounded to "nearly three-quarters," the underlying claim still needs to trace back to a real, current, correctly described source. Rounding changes the phrasing, not the verification obligation.
Is it ever acceptable to publish a statistic I could not fully verify?
Only if you say so plainly, using language that reflects genuine uncertainty, such as attributing it to where you found it and noting you could not independently confirm the original data. Presenting an unverified number with the same confidence as a verified one is the exact failure this whole process exists to prevent.