Everyone's Learning to Prompt. Nobody's Learning to Doubt.
Written by Gavin Dixon - Director of Global Perspectives
Give an experienced person a piece of AI-generated work in their own field and you'll often see them slow down somewhere in the middle of it. Not because the writing is poor. It reads perfectly well. It's what the writing is saying that doesn't hold up.
Now ask them how they knew.
Most can't tell you, or not precisely. It felt thin. Or they've seen something like it go wrong before and it didn't sit right. Whatever it is, it's years of pattern recognition that nobody ever wrote down, and it's doing more work in your organisation than anyone has accounted for.
There's no shortage of courses teaching people to write better prompts. There are far fewer teaching people to tell when the answer is rubbish.
Which raises the question of where that judgement came from. Nobody acquired it by attending a session on it. They got it by producing a lot of mediocre output and having someone more experienced explain what was wrong with it. The first drafts nobody sees. The documents that come back covered in red ink. Most of that work was thrown away afterwards.
All of which is the category of work AI now does well. That's the part I don't think organisations have followed through yet.
The thing at risk here is subtler than checking the output.
AI writes with exactly the same assurance whether what sits underneath is sound or entirely invented. Nothing on the surface tells you which one you're dealing with.
Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers about 936 real examples of AI-assisted work, published at CHI 2025. What they found is that people stop thinking critically when they lack the skills to inspect and guide what the AI hands them. The people who have low scrutiny skills question the output least and feel fine about it.
So it lands on managers. Passing this on means examining how you know what you know, which most experienced people have never had to do. It means being willing to say "something's wrong here and I can't yet tell you what" in front of a junior. In Japan that's a heavier ask than it sounds, for reasons most people reading this will recognise.
Then it has to be said in a way that actually has impact. "Not rigorous enough" transfers nothing. "You've accepted the claim in the third paragraph without asking what it rests on" transfers something. That's harder with AI output than with a colleague's draft, because there's no reasoning behind it to point at. You're building the explanation out of your own head every time, pitched at what the other person already knows.
And if your team works across two languages, as plenty of teams here do, it gets harder again. Precision is the entire value of the explanation, and precision is the first thing to go when one of you is working in a second language.
And it means giving up on "they'll pick it up the way I did." That one's forgivable. It used to be true.
Something to try this week. When you kill a piece of AI output, don't quietly fix it. Say out loud what made you stop. That's the whole transfer, and it costs about forty seconds.
If your three most experienced people left next year, who would still be able to tell that the AI had got it wrong?
Most organisations can't answer that, because nobody has ever measured it.
And the habit above only works one manager at a time. Getting it to hold across a management layer is a design problem rather than a good intention, and that's the part we work on at Global Perspectives.
