What AI can’t do: the editorial decisions that still require a human in the room

“Keep a human in the loop.”

It has become the reassuring line. After the failure stories and the warnings, it sounds like a solution.

But a loop with a human in it is, in itself, not a safeguard. Not unless someone can say what the human is there to catch.

In many organisations, “use your judgement” is not an instruction. It is a shrug. It sounds like rigour but asks for nothing specific. The draft still goes out. The person who waved it through often cannot tell you what they checked, because no one told them what to check.

That is the gap. Not whether a human reviews AI-assisted work, but whether the review knows what it is looking for.

A human is there to catch the claim that the source material does not support. The caveat that was true in the original and disappeared in the rewrite. The confidence that the evidence has not earned. The thing the organisation has chosen not to say, which the model cannot know and may say anyway. Fluently.

None of that shows up as an error. The draft reads well. It passes a spellcheck, a tone check, and a quick read from someone senior. It fails only against the question nobody asked: is this accurate enough, careful enough and defensible enough to leave the building with our name on it?

“Human judgement” is not the answer to AI in communications. It is the phrase that lets us feel covered while the actual work goes unnamed.
The work is naming what the human is there to catch. And then making sure they do.

That is the discipline – and it can be taught. It’s what we do at Word of Art.