How do I stop AI drafts from inventing statistics?
Two moves. Never ask for a number the model has no source for, and make every number in a draft traceable to an input you supplied. Generation invents specifics because specifics make prose sound authoritative, and a model asked to write persuasively will supply them. The fix is structural: numbers come from your systems as inputs, and any figure appearing without a source is treated as a defect, not a style note.
Why invented specifics appear at all
Confident prose contains numbers. A model trained to produce text that reads like expert writing will produce text with the surface features of expert writing, and a plausible percentage is one of those features. It is not lying in any useful sense; it is completing a pattern.
Which is why prompt instructions like "do not make things up" underperform. They ask the system to override a pattern rather than removing the situation that produces it.
The rule that removes most of them
Numbers are inputs, not outputs. If a figure belongs in a piece, it is retrieved from a system that knows it — your job records, your call data, your ad platform — and handed to the drafting step as a fact to place. The model arranges; it does not source.
That is the difference between generative writing and grounded writing, and it is the same principle that keeps AI-written business reporting trustworthy. A summary that quotes a number it was given is verifiable. A summary that recalls a number is not.
A review pass that actually finds them
Do not read for accuracy in general. Scan for specific shapes, which is much faster and much more reliable.
- Any percentage. Trace it or cut it.
- Any named study, survey or report. These are the most commonly fabricated element in AI drafts.
- Any count. Customers, locations, years, integrations, languages.
- Any superlative. Most, fastest, only, leading. Each one is a claim someone can dispute.
- Any version number or date. Confidently stated and frequently wrong.
Categories to prohibit outright
Some claims should never come out of a generator regardless of review, because the cost of one slipping through is disproportionate: client results and case metrics, competitor claims, certification or partner status with any platform, and legal or regulatory thresholds. Put these on a banned list enforced before review rather than during it.
The general principle is that a content pipeline should not be capable of producing a claim your business cannot stand behind. Designing that constraint into the workflow is part of building it as a system — see content automation and the model-side controls in AI model integration.
Topics: hallucination · fact checking · grounding · editorial review
Have a version of this question about your own business?
The useful answer usually depends on which systems you run and how they're connected. That's a conversation, not a blog post.