The claim, in one paragraph
Production stopped being the bottleneck. A machine will now draft, summarise, model, translate, design and code faster than the person who used to. What it will not do is decide what deserves doing, judge whether what came back is any good, or answer for the result. Those were always the scarce parts and they were hidden inside the production, which is how most organisations came to believe they were buying speed. They were buying speed and spending judgement, and only one of those appears in the business case.
What judgement is, stated narrowly enough to be useful
The word is used to mean everything from taste to ethics to experience, which makes it impossible to manage. The narrow version: judgement is deciding well under conditions where the right answer is not recoverable from the information available. If the answer could be looked up, computed or proceduralised, the task needed knowledge or skill rather than judgement. What judgement is, and how decision quality differs from outcome quality, which is the distinction most organisations never draw.
The consequence of the narrow definition: judgement cannot be taught by explanation. It is built by making decisions of the kind in question and finding out how they turned out. That is the reason delegation has a cost, and the reason the cost is invisible until it is large.
It is being removed rather than replaced
This is the part I argue hardest. It is where I part company with both the enthusiasts and the doom side. The machine is not taking over your judgement. It is removing the occasions on which yours would have been exercised, which produces the same result more slowly and with nobody deciding to do it.
What disappears first is the demand. Nobody announces that a class of decision has stopped being made by people. The work still gets done and the output often improves, so the loss shows up in no report. It appears years later as a capability the organisation no longer has, in a moment when the system cannot help. Does AI weaken human judgement is the evidence reading of that claim, and capability debt is the name for the bill.
Which makes it a governance question, not a training one
If judgement were a skill, the answer would be a curriculum and the problem would belong to the learning function. Because it is built by exposure to real decisions, the answer is a set of choices about which decisions people keep making, and those choices are made by whoever designs the work. That puts it on a board agenda rather than in a development plan.
Five decisions follow, and they are the spine of the board work on this site: which decisions machines may make in the company’s name, who can stop each system, what people must remain able to do unaided, how anyone would find out if something went wrong, and what management’s assurances actually rest on. Board oversight of AI sets them out against what a UK board now has to declare.
Where I disagree with the consensus
Against the optimists: that a better output is evidence of a better process. It frequently is not, and an organisation measuring only output will not see the thing it is losing until the measurement it needed was the one it stopped taking.
Against the pessimists: that this is deskilling, arriving on its own, to be resisted. It is a design problem, it is decided by people, and most of the decisions are still open. Drift versus design.
Against the middle: that keeping a human in the loop is the answer. A person who approves at machine speed, without the standing or the time to disagree, is a signature rather than a safeguard. Human in the loop is not a safeguard.
What would change my mind
Stating this is part of the method. The argument weakens considerably if the capability loss turns out to be recoverable at will, so evidence that lapsed judgement returns quickly once demand returns would matter: can you regain a skill you have lost. It weakens again if oversight arrangements that look ceremonial turn out to catch errors at a decent rate in practice, which nobody has measured well.
And the strongest version of the case against me is worth stating plainly: every general-purpose technology has removed a class of human judgement, few of those losses were mourned for long, and calculators did not ruin arithmetic in the way that was predicted. I think this one is different because of where it lands rather than because it is bigger, and that is an argument rather than a finding.
Where the evidence sits, and why it is somewhere else
Every claim above that can be tested has studies behind it, graded, dated, and annotated with what each one does not support. They live at the evidence base, and the questions this work is organised around are at the questions.
The separation is deliberate. This page is one person’s argument and is labelled as such. The evidence base is a secondary reading of other people’s work and is labelled as that. A reader who wants to disagree with me can do it from my own sources, which is the arrangement I would want from anybody asking me to believe them.
