A junior analyst sent me a competitor breakdown last month, built with AI, and it was beautiful. Clean structure, confident headings, the right shape on the screen. I read it twice while she waited on the call, and the second time the floor went out from under it: the pricing was a year stale, and the headline recommendation was something the client had tried and shelved months before. She had no idea any of it was wrong. She could not have known. Nobody had ever made her find out the hard way.
That is the real shape of human judgement in the age of AI, and it has a paper now. David Duncan, writing in Harvard Business Review in February 2026, watched a consulting partner use generative AI brilliantly while his juniors used the same tools and could not tell good work from bad.1 The partner had the calibration. The juniors did not. The tools could not hand it over.
Judgement is not taught. It accumulates, slowly, through being wrong in front of people who matter. The analyst gets the deck torn apart. The junior lawyer watches a clause bleed red ink. The young operator writes a brief that goes nowhere and, six months on, understands why. That is the formative work, the consequential repetition that builds the earned instinct to look at a deliverable and know it is off before anyone else does. AI does that work now. So the person who never does it never builds the thing that lets them judge it.
McKinsey’s 2026 work on skills makes the trade plain.2 Negotiation, leadership and problem-solving sit at low exposure to automation and rising value. Detail orientation, quality assurance and document review sit at high exposure, handed to the machine. The skills climbing in worth are the ones you only build by doing the skills being automated away. Read that twice. It is the whole problem in a sentence.
Researchers at Johns Hopkins put it without softening in early 2026: junior workers cannot manage the AI because they lack the experience the AI was meant to spare them.3 The bottom rung of the ladder is gone. Nobody has built the replacement. We are mid-fall and still calling it efficiency.
Why human judgement in the age of AI is a leadership problem
This sits at the top of the organisation, not the bottom. A manager who never did the underlying work, leading a team that never did the underlying work, signs off on AI output with nothing under the signature. The analysis looks sound. The argument coheres. No inner voice says wait, that is not right, because that voice is built from a hundred times you were wrong and found out. Ethan Mollick’s consultant research shows the upside is genuine: people above the performance threshold gained 17 percent, those below it 43 percent.4 Everyone improved. But the floor under them was poured years earlier, by hand. Who is pouring it for the people starting now?
The gap, made visible
Try this. Take the most junior person in any function and ask them to read a piece of AI-generated analysis and say where the reasoning is weak. Ask whether the argument holds, where the gaps are, what it leaves out, rather than whether it looks finished. If they can name it, you have someone building judgement. If they cannot, you have a team that produces work it cannot evaluate, which is a different and harder problem, and one most leadership agendas have not reached, because the output still looks fine.
Protecting the training ground
Three questions are worth sitting with, for any team leaning hard on these tools:
- Where does real consequence still reach people early on? Feedback on whether the substance was right, and why it was wrong, not a note on how well they drove the tool. Review only the output and the lesson learned is that the output passed.
- Where do people still get to be wrong and have to find the fault themselves? Being told it is wrong is a world away from hunting down why. AI shrinks the obvious errors and leaves the subtle ones, which are harder to see and costlier to miss.
- Where are you building stretch without instant correction underneath? Real stakeholder exposure, case review with the tools switched off for the first pass, a decision made before the machine gets a vote. Judgement grows in managed uncertainty, or it does not grow at all.
What to do this week
Pick one junior person. Ask them to assess a piece of AI-generated work in writing, with no AI to help, in an hour. Form your own view first, then read theirs against it. What you learn will tell you more about your leadership pipeline than the last three performance reviews put together.
The productivity is real. The amplified senior people are real. So is the slower thing underneath: organisations getting very good at producing work, and thinner every quarter on the people who can tell when it is wrong. That skill can still be built in humans. Whether anyone is choosing to build it is the question we are not asking out loud, and the one I keep turning over after the call ends and the beautiful, wrong document is still open on my screen.
Sources:- David Duncan, “How Do Workers Develop Good Judgment in the AI Era?”, Harvard Business Review, February 2026[↩]
- McKinsey Global Institute, “Agents, robots, and us: Skill partnerships in the age of AI”, 2026[↩]
- Johns Hopkins University, “Will AI make human workers obsolete?”, 23 February 2026[↩]
- Ethan Mollick / Valence, “AI agents, agentic work and the future of work”, 2026[↩]
