BACKGROUND
This project was a human-in-the-loop example documenting six real errors caught during the review of AI-assisted content drafts across three different pieces and content types. The goal was to show what AI errors actually look like, how each was identified, and what would have happened if it had been published.
I documented each error with its type, the exact AI output, a diagnosis of what was wrong with it, the fix applied, the review step that caught it, and the realistic downstream consequence if it had gone live. The errors ranged from a fabricated WHO statistic and an incorrect legal threshold for non-standard working patterns, to a tonal framing that would have alienated the target audience, a competitor naming policy violation, and a passage that drifted from legal information into legal advice.
The result was a document that made the case for structured AI review as a professional skill requiring domain expertise, not an administrative checkbox, because most of the errors required subject matter knowledge to catch rather than simply careful reading.
