How do you stop AI from making things up in published content?

Fluency is not evidence of accuracy

The intuition that wrong content sounds wrong comes from an era when wrong content was written by people who did not know the subject and it showed. Model output does not show. A fabricated code requirement and a real one are written in the same register.

This has a direct process consequence: review must compare against something external. A reviewer reading a draft and asking "does this seem right" will approve a confident error, and will do so faster than they would have approved a hesitant truth.

Three controls that actually work

• Closed-set facts. The drafting step receives an explicit list of facts it may use and is instructed to use nothing else. Anything it needs and does not have becomes a flagged gap for a human, not an invention.
• The no-number rule. No numeral appears in output unless it was in the supplied source or is definitional, like a formula. This one rule eliminates the largest share of dangerous errors, because the costly fabrications are almost always quantities.
• Claim extraction and check. A separate pass lists every factual assertion in the draft as a flat list, and each is checked against source. Reviewing twelve extracted claims is faster and far more reliable than re-reading eight hundred words of prose.

Where a fabrication costs the most

Rank your risk categories and put the heaviest review where the damage is real: code and permitting requirements, warranty and manufacturer terms, licensing and insurance statements, anything about safety, claims about competitors, and claims about your own capabilities and coverage.

That last one is underrated. A page confidently describing a service you do not offer generates calls you cannot fulfill and erodes trust at the exact moment a customer was ready. It is also the category models get wrong most often, because they fill gaps in your business description from context. Supplying an explicit, maintained description of what you do and where — the same record that drives the rest of your content system — removes most of that category.

Design the pipeline so review is cheap

If verification is a person opening a document and reading carefully, it will be skipped under deadline pressure. If it is a short list of extracted claims with sources beside them and a checkbox each, it survives.

The same principle governs any operational AI workflow: the human step has to be small, specific and fast, or the system quietly degrades into an unreviewed pipeline that still has a reviewer's name on it. Building the check into the tooling rather than into the process document is what makes the difference.