We almost published a stat our AI made up, now every claim gets checked before it goes out

Had a near miss last week worth sharing. We were drafting a carousel for a client and used an AI assistant to pull a few supporting stats. One of them was a great number, very specific, very quotable, attributed to a named industry report. When I went to source it for the caption, the report did not exist. Not misremembered, not the wrong year, invented, with a real sounding org name bolted onto it. We caught it because we source every claim now, but a year ago that number probably goes out in the post.

This is a content problem before it is an AI problem. A made up stat in a brand's post is a credibility event, not a small oops. Someone in the replies always fact checks, and your viral post cites a study that does not exist is the kind of ratio that trails an account for months. With AI fabrication clearly rising, one widely shared report found fake references in published research climbing fast over the last couple of years, the base rate of plausible but false claims in anything AI touched is going up, and our content sits directly downstream of that.

The workflow we landed on is not clever, it is just enforced. Every factual claim in a post has to trace to a real, linkable source before it ships, the same standard a journalist would use. For digging claims up and pressure testing them we lean on research tools that show their sources rather than just handing back a confident sentence. One we have used is apodex, mostly because it flags when sources disagree instead of smoothing them into one tidy number, which is exactly the moment you want to slow down. The specific tool matters less than the rule though, nothing factual goes out unsourced.

If you run brand accounts, the takeaway is that the reputational downside of one fabricated stat dwarfs the few minutes it takes to check it. Build the verification step into your content calendar the same way you build in approvals and scheduling. The brands that get publicly embarrassed this year will mostly be the ones that treated AI output as finished research instead of a confident first draft that still needs a human to verify every number in it.