So here we have deskilling, as people call it these days — a bit of a nuisance. It is another observation from the October “AI Transformation of Healthcare” conference. This time, the fear that we will forget how to think if we just click through AI results.
I would probably write myself off socially if I claimed that the concern was unfounded. I am just convinced that it misses the point.
The problem is the volume of work, not the loss of skill
The problem is the volume of routine work we cannot get through at a decent quality. Checking, comparing, judging similar cases over and over. If I can no longer simplify those blasted processes and strip out the steps that can be stripped out, then why not bring in AI — which, this time, is not the thing being cursed?
And what is worse, what if the repetitive activity is a key part of the process — an imaging examination, for instance? Rows of images, similar tasks, attention that has to last the whole shift. Is the problem really that, with the help of AI, I start to lose my skills? Or is it rather that our brain does not deliver the machine-like stability of performance?
Human performance fluctuates. We get tired, we lose focus, towards the end of the day we see things differently than at the start. That is not a moral failure; it is a property of the brain. A machine, in this sense, does not get bored and does not wander. I am not saying it is infallible. I am saying that its error rate can be described, and the human one usually cannot. That is why it seems a shame to me to fear the tool before we know how a person performs at the same work.
Enough questions. The whole thing comes down to this: with AI, we can measure the reliability of the output. And when we do it properly, we know, transparently and repeatably, how it performs. Fine, we can debate the quality of the dataset, what the model was tested on and how far that carries over into practice. About a person, we do not know. We ask about education, experience and the impression that they are “good”. That is not the same kind of information as reliability measured over and over. We only guess, believe, hope or fear — but we usually do not quantify the quality of human output.
When we then put AI, for which we have a number, next to a person, for whom we do not, the person can look worse even if they have not changed. So if it seems somewhere that the quality of human output falls when AI is used, the problem will not be, in principle, with the AI. It is more that there was simply no measure of the quality of human output. AI merely makes visible what we have been living with, not how we have got worse.
When human work really is measured
It is no accident that the debate so often turns on imaging. Repetitive reading is the core of the work there — and at the same time one of the few places where human performance can be compared with a number at all. A typical example is colonoscopy. The indicator is the adenoma detection rate: the share of examinations in which the doctor finds at least one adenoma. Precisely because that number exists, it was possible to find something out at all when, in August 2025, an observational study from Poland was published in Lancet Gastroenterology & Hepatology.
Nineteen doctors at four centres were compared over three months before a polyp-detection system was introduced and three months after it — always in examinations without AI. The question, then, was not how a person performs with AI, but how the same doctor performs without it after growing used to it in the rest of their work. The adenoma detection rate fell from 28.4% to 22.4%. That is the result now cited as evidence of deskilling.
I would be cautious about it, though. It is an observational comparison, not a randomized study. The patient groups differed, and examination time, which strongly affects detection, was not tracked. The number of adenomas per examination, and the detection of advanced findings and of cancer, did not change statistically. The authors themselves treat it as a hypothesis to be tested, and call for a proper prospective study.
And above all: what matters for the patient is how well the person and the AI do it together. There, randomized studies repeatedly show that detection rises. A meta-analysis of 28 randomized studies with nearly 24,000 patients, published in 2024 in Gastrointestinal Endoscopy, reports that with AI the adenoma detection rate was 20% higher — mainly for smaller lesions; for advanced adenomas, no difference appeared. If we are afraid that skill will dull without AI, that is an argument for measuring people regularly, not an argument against a tool that improves detection when they work together.
A measure does not have to be confined to healthcare. Where the output can be compared with the facts, or with a second reading, human reliability can be tracked in the same way as a machine’s: on a sample, repeatedly, not once in an audit. Without that, deskilling is easy to talk about and hard to judge. Not so that we stop trusting people, but so that we do not have to argue it out from our desks. First the number, then the conclusion: whether we really are getting worse, or whether we are only now seeing what has been there all along.
Do you measure at all, where you are, how reliable human output is? Or do we only guess and hope?
I published a shorter version on LinkedIn on 3 November 2026.