The judgement that cannot be checked

Posted on: 23 July 2026

Read enough job descriptions written by people who make their living from selection and a pattern becomes hard to ignore: almost none of them describes the work being hired for. They describe a wish, and usually an incoherent one, ten years of experience with a tool that has existed for five, a degree demanded for tasks that never come near it, and beneath that the familiar list of personal qualities lifted wholesale from the previous advert. The first concrete act of the trade, performed long before anyone walks into a room for an interview, is this document. And it is already here that the process loses the right candidates, because it turns them away before it has seen them.

I raise it because the season has produced another of those posts that professionals write in defence of their category the moment the category feels the pressure at its back. The argument is familiar and reassuring. Artificial intelligence processes data, the human being interprets people; the algorithm accelerates, but it is the expert eye that confers meaning, and that meaning the machine cannot touch. Empathy, intuition, the ability to read what the candidate does not say. It closes on a rhetorical question, which competencies will AI never be able to read in a candidate.

That is precisely the wrong question, and it is worth understanding why. The competency the recruiter holds up as a bulwark, the fine reading of a person, is by its nature the least verifiable thing there is. Nobody outside the head of the person who decided knows why one candidate was preferred to another. There is no written criterion, no measurable error to fasten onto, no way to go back and say: here you were wrong, and for this precise reason. Intuition leaves no traces that can be contested. And it is exactly this absence of traces, not the empathy, that is being defended.

A selection system, however crude and however loaded with biases of its own, carries a flaw that for the category is fatal. It can be opened. It can be called to account. Its error can be measured and its criterion corrected. It is inspectable. Human judgement is not, and its survival strength lies entirely there, in never having to be put to the test. When you defend the irreplaceability of the expert eye, what you are defending is your own unfalsifiability.

Not every profession has reacted this way, and the difference is instructive. When diagnostic AI began to beat radiologists on specific image recognition tasks, radiologists did not retreat behind clinical instinct. They claimed the measurable ground: responsibility for ambiguous cases, documented accuracy where they had it, the integration of the reading into an actual clinical history. They defended the part of the work that can be verified, and on that ground they held. Whoever defends the verifiable survives verification. Whoever entrenches on the invisible protects only the fact that no one will ever be able to prove them wrong.

So why choose the worse trench? Because the right one is unflattering. The real value of a good recruiter is almost never interview telepathy. It is access to a network and the willingness to put one's face on a risky choice, so that if the hire fails there is someone to blame who is not the chief executive. It is reputational cover and the intermediation of trust, valuable and defensible work. Admitting as much, though, means declaring oneself an insurer rather than a reader of souls, and the insurer has rates, not an aura.

I watched this happen from inside a trade I once knew well. When digital arrived in film post-production, a portion of the professionals who lived by the eye, by a sensitivity to colour built over years of celluloid, responded by explaining that the machine would never have that taste. Some of them were right and are still working today. Others were merely defending the fact that their taste could not be measured, and when the tool made a good part of it measurable, the genuinely irreducible fraction turned out to be a great deal smaller than they had sworn.

There is a more uncomfortable reason still. For the better part of a century the research literature, going back to Paul Meehl in the 1950s, has shown that a banal checklist applied with consistency outperforms expert judgement in predicting almost anything that turns on people, from criminal reoffending to academic attainment. The finding is seventy years old. The astonishment and the refusal are new every time, because that research says the one thing no professional of judgement wants to hear: the prized part of your work, the instinct, is the easiest to replicate with a dull and steady method. It is not that the experts are unaware of the studies. It is that the studies are about them.

Here it helps to clear the moral reading out of the way, because it does not hold. The recruiter does not keep the lights off out of cunning; the lights stay off because the client wants them off too. Whoever commissions a search buys, along with the candidate, the possibility of not being the only one held responsible when the choice turns out badly. The opacity of the criterion serves the one who decides just as much as the one who executes. It is a two-sided equilibrium in which neither party has any interest in setting down in black and white why this person and not that one. An explicit system threatens both. That is why the resistance is so broad and so poorly argued. It is not defending a technique, it is defending a shared convenience.

Recruitment is only the most naked case, because it produces that embarrassing artefact, the job description, and leaves it lying about as evidence. The mechanism, though, is present wherever a judgement certifies itself. Consumer credit, before scoring made explicit and therefore contestable who was creditworthy, lived in the same half-light: the local bank manager who "knew his customers". A good part of consultancy sells experience when what it is really selling is someone with whom to share the responsibility for a decision the client had already taken. Making the criterion visible, in all these cases, does not remove competence. It removes shelter.

One can object, and it is the serious objection, that the opacity is not a defended rent but merely a lag: selection remains a matter of the eye because until yesterday there were no better instruments, not because anyone deliberately keeps the lights off. The objection deserves respect, and there is only one way to settle it, which is to watch what the category does now that the instruments are arriving. If it adopts them, tests them, uses them to make its own choices more legible and more correctable, then it was a lag in good faith. If instead the prevailing reaction is the article explaining why the human eye must not be touched, the category is answering the question for itself.

Which is why the question in the post, what AI will never be able to read in candidates, is not the one to ask. The real question, the one no one in the category has any wish to put aloud, is a different one. What have we never wanted anyone to read in our decisions.