AI content that does not make things up

AI can write a thousand articles a week. The problem is that some of them will contain a number that is slightly wrong, a quote nobody said, or a confident sentence about something that never happened. On a site that earns from trust, one of those is enough to hurt.

On a recent engagement we published 430 data-led articles in two weeks, in two languages. Here is how we kept them honest.

Separate the facts from the writing

The AI does not decide any figure. Every number is computed by script from the source data first, and handed to the model as a fixed input. The model’s job is to explain and structure, not to remember or estimate. Charts are drawn from the same numbers, so text and chart cannot disagree.

Check before it goes live, automatically

• Numbers: every figure in the draft is matched against the computed data. A mismatch blocks the article.
• Quotes: every quotation is matched against its source text. No match, no publish.

Thin pages were part of the problem too. Before the engagement, one flagship series had 16 of 16 pages at 43–47 words. They were rewritten at 1,235–1,475 words, every figure from data.

Keep humans where judgment matters

People stay responsible for what gets published. The AI makes the volume possible; it is not the editor-in-chief. We also audited 300 existing section pages (4,575 findings) and rewrote 255 of them, because new content on top of a weak base wastes the effort.

Publish when it can still rank

Speed is only useful if it lands at the right moment. Before the engagement, articles regularly went live too late to rank for the searches they targeted. Volume without timing is wasted, so when to publish is now something we test, not something we guess.