A young doctor sits in front of a patient with vague symptoms. In the system before her, an algorithm has already sorted the medical record data, suggested differential diagnoses, and ranked probabilities. The system is probably doing something right: it processes more data faster than she could have done alone, and it does not overlook the blood test value that she might have dismissed. But something else is also happening, something that concerns her and not the patient: the questions she does not need to ask are precisely the questions she would have learned something from asking, and they are also the questions that could have led her to conclusions the data did not already contain.

A healthcare system under pressure

It is worth starting with the situation as it actually is. Norwegian hospitals operate under pressure that has been increasing for years: growing waiting lists, personnel shortages, and administrative burdens that eat into clinical time. In this situation, it is not only legitimate, but necessary to adopt technology that can shorten waiting lists, reduce error margins in image interpretation, ease administrative burdens, and distribute resources more precisely. Where routine tasks take time away from clinical presence, automation can free up space for what actually requires a human. No one is served by romanticizing the inefficiency of a system that is already failing patients, and technological resistance should not become a hidden conservatism.

The argument that follows is therefore not that today's system is good enough and should be protected from change, but that in the rush to repair what is broken, we risk destroying something else without even noticing.

When decision support shapes attention

In clinical encounters, much of what makes healthcare good is difficult to capture in indicators: the continuity of a relationship over time, the trust that makes the patient say what she is actually feeling, the doctor's ability to read something in the situation that is not documented anywhere. These things are real, they have clinical significance, and they are systematically underrepresented in the databases on which systems are trained. This is not because they are unimportant, but because they are difficult to measure, and what is difficult to measure rarely ends up in training data.

The more interesting point, however, is not that the unmeasurable is underrepresented, as this is nothing new. The new aspect is the mechanism: as decision-support systems gradually shape what clinicians focus their attention on, the clinical culture shifts. The systems direct attention to what they are built to see, and what they are not built to see recedes into the background. No one has decided that the measurable is most important, but the infrastructure has done so, because attention is a limited resource, and it is guided to where the tools direct it.

Clinical judgment as something cultivated

Clinical judgment is a capacity formed through practice, not an inherent trait. The ability to ask the right question at the right time, to feel unease at a vague symptom description and let that unease lead to reading what is left unsaid in a consultation room: this ability is the result of years of accumulated experience, of having made mistakes and understood why, of mentorship in a broad sense. The central question becomes not just what is overlooked in the moment, but what is eroded over time. 

Craftsmanship is the obvious parallel. What is passed from master to apprentice is not technique alone, but a way of seeing and sensing the material. This way of seeing cannot be taught directly; it is formed when the apprentice must try, fail, experience uncertainty, and find the way forward. It is this process—standing in what one does not understand long enough for understanding to emerge—that is the productive friction. 

In most craft traditions, this has been understood: the introduction of tools that remove friction has been managed with caution, because friction itself is part of the training. Medical education retains cadaver dissection despite advanced simulation tools. The point is not to reject the tools, but to design the learning pathway around them, with an awareness of what is lost when the process changes.

The double loss

Within this lie structural, and not just technical, challenges. As AI systems take over more and more of the diagnostic work, a double loss occurs. The first is the loss in the moment: the dimension of the clinical encounter that the system does not see. The second is the loss over time: that the conditions that shape judgment gradually disappear. A new generation of clinicians develops capacities in response to the tasks they actually solve. When the nature of the tasks changes, their development changes with them. The loss is significant, yet omitted from the ledger. And what makes the loss difficult to detect is that the resulting workflow is not easily distinguished from the workflow produced under previous conditions. A clinician shaped by decision support appears competent within the parameters she operates in. It is only when the situation deviates from the norm, for instance, when the system fails or when the situation is genuinely new, that the difference in development becomes apparent.

It is easy to object that automating diagnostic work will free up the doctor to spend more time on precisely those relational, judgment-based tasks. The objection has some merit, and the promise of freed-up time is real, but it rests on two assumptions that are rarely made explicit. The first is that the tasks being removed were merely time-consuming and not simultaneously formative—that the doctor learned nothing from asking the questions herself. The second is that the attention formed when interacting with a system that has already evaluated and sorted is of a different nature than the attention formed without such a system. A clinician who grows up with decision support is trained to evaluate suggestions rather than make evaluations herself, and this shapes a different kind of clinical gaze: less exploratory, perhaps less willing to ask the question the system has not asked. A clinician who has time to spare because the system has asked the questions for her is not the same clinician shaped by having asked the questions herself. 

What kind of clinicians we want to shape

The question of how much should be automated is therefore not just a technical question about the precision of the systems, but a question of what healthcare is for. If the answer includes that patients should be met as whole human beings by clinicians who carry responsibility for the assessments they make, then this commits us to certain capacities that must be formed and maintained, and certain conditions that must be protected so that they can be formed.

This means that decisions about what to automate are more than technical purchasing decisions. They are decisions about what kind of professionals we want ten, twenty, and thirty years from now, and what kind of profession they will belong to. These decisions require a conversation among more parties than just those who build and sell the technology: it belongs to educational institutions, healthcare trusts, professional associations, and public debate. But those who build the tools also have a responsibility to think beyond the next product version, because the tools they design shape the learning conditions of those who will use them.

The young doctor in front of the screen does not have to choose between the system and her own clinical judgment. But someone—her educational institution, the hospital she works at, the vendor that built the system—must have thought through where her productive friction will come from as more and more of it is removed. The goal of automation is correct: to free up time and attention for what requires a human. But it requires us, for every task in every domain, to ask not only if this can be done by a machine, but also what the human becomes by doing it themselves, and whether it is inconsequential that humans are no longer shaped by this process.

Read an extended version of the article here