As a leader, every piece of work we receive tells us two things: the substance of the document, and something about its author - their grasp of the subject, their ability to draw conclusions, their attention to detail.
A CEO of an insurance company brought this home to me recently. New in the job, he described a weekly management report that came to him. It was long, and often not very good. I asked whether it was weak because the author didn't understand the substance, or only because they couldn't present it well. The answer was "a bit of both".
Naturally enough, I asked whether he'd considered using AI to produce the reports. It would save time and lift the quality, which seemed like a pretty clear win. He had. But it raised a real concern: he'd lose the signal the raw work gave him about the capability and expertise of his people.
Like mine, your LinkedIn feed is probably awash with posts about the importance of judgement in the AI age (as if it didn't matter much before, but I digress). The real question isn't whether judgement matters but who has it, and how we tell who has the expertise to make which calls - just as AI removes one of the few signals we had to go on. Another related question is how that expertise, the foundation of sound judgement, gets built in the first place - and how AI is changing when and where that happens.
Both deserve more than the feed gives them. Producing that raw work was never only about the output. Working a problem through - reasoning toward an answer, getting it wrong, seeing why - is how the judgement gets built. Expertise forms in an uncomfortable way: by doing the work badly until you can do it well. AI now does the work, and that changes two things at once. Your job shifts from producing it to checking it, and checking it well takes the very expertise you'd once have built by doing it. At the same time, producing it was where that expertise came from, so as the work moves to the machine, the building moves with it. That second half is well-worn ground: the collapse of the junior pipeline, the entry-level rungs that used to turn graduates into seniors. The first half gets less attention, and together they make a bind - we're asked to verify more and more with judgement we're building less and less.
The obvious reply is that people just need to use AI well. That's true, and it hides a problem. Using it well means judging what it hands you - directing it, catching where it's wrong, noticing what it left out - and that isn't a skill you lay on top of expertise; it is the expertise. Having it doesn't mean the tool speeds you up: in one controlled trial, experienced developers were 19% slower with AI on code they knew well, while being sure it had sped them up. What expertise buys you isn't a boost, it's the ability to tell when the tool is wrong and when it's just in the way. Without it, you can't tell a sound answer from a confident wrong one, and the tool doesn't lift you, it just makes your not-knowing look finished. There's a name for what's going on underneath, the expertise reversal effect: the support that helps a beginner is close to the opposite of what an expert needs. Handing the work to AI is the expert's move - useful if you already hold the judgement to check it, corrosive if you're using it to skip the building.
The trouble is that we treat "beginner" as a stage you pass through once and leave behind. It isn't - it recurs. Every new domain drops even the most senior person back to the foot of a curve, and a leader is a permanent novice across most of what they oversee, AI included. We're also poor at telling which one we are at any given moment. So the trap isn't that you can't judge your own level. It's that, not realising you're a novice, you reach for the tool that does the work - at the very moment you needed to do the work yourself.
And there's a human piece the discourse skips. If we're poor at reading our own level, someone else used to do it for us - and the way a good manager always knew was by watching the work. The rough early attempts showed them where you still needed a hand and where you'd earned room to run. That's the signal the CEO feared losing, seen from the other side. Strip out the raw work and you don't only stop people building judgement, you blind the person whose job was to know how much help each of them needed, and when to stop giving it.
Follow all this forward a few years, and the two questions from the start collapse into one. The person who let AI do the thinking didn't build the judgement, and because they didn't, they can't tell when it's missing - in their own work or in anyone else's. You get people who can neither produce sound work nor recognise it when it lands on their desk, and a layer above them that often can't tell either, because everything still arrives looking finished. The gap doesn't announce itself. It grows quietly, until something that matters rests on judgement that was never built.
So the work is to put the formation back in on purpose, because it won't happen on its own. Two moves follow.
Your own judgement about AI is a novice's judgement, however senior you are elsewhere. It builds the only way it can - by doing the work yourself, on real problems, before you decide anything about it rather than after. And the people around you will build judgement only if the work is designed so they must. Have them reason the problem through first, then use AI to test what they landed on, not the other way round. It's slower, and it's the version that makes the person.
There's a more hopeful version of this, and I've started to see it in my own work. Left to its defaults, AI strips out the feedback that used to build judgement. Aimed deliberately, it can do the reverse. It can show you where you're genuinely expert and where you're only guessing - the thing we're worst at seeing in ourselves - and it can shorten the distance between doing something and learning whether it was any good. That is the feedback loop expertise has always needed, run faster. It isn't what the tool does on its own. But it is within the gift of whoever designs how the work gets done, and it is the difference between AI that quietly hollows judgement out and AI that helps build it.
Sources
METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (2025, 16 developers, 246 tasks)