Shadow AI is not a physician phenomenon

Shadow AI is when staff use AI services the organisation has not approved, does not control, or does not even know about - ChatGPT in a private browser, for instance. The term is borrowed from "shadow IT", and a recent study shows that 357 Swedish physicians work this way today.

The study is about physicians and says nothing about psychologists, nurses or counsellors. But I find it hard to believe the phenomenon stops with doctors. The drivers exist across clinical work: heavy workloads, large volumes of information, and situations where complex material must be weighed quickly.

Picture a psychologist in the middle of a neuropsychiatric assessment. Case history, self-report questionnaires, informant reports, observations and test results are on the desk. The differential diagnosis is hard, and the workplace has no approved AI support.

But at home, in the browser, there is a general-purpose AI service. The psychologist pastes in parts of the material and asks:

Which differential diagnoses should I consider? Which findings support or contradict ADHD? Have I missed any contradictions?

The answer arrives in seconds - sorted, structured, perhaps with an impressive piece of clinical reasoning. For the psychologist, the value can be enormous.

At the same time, questions arise that no one has answered. What information has left the organisation, and where is it processed? Is the tool fit for clinical judgement at all? How much does the AI's reasoning colour the psychologist's own conclusion - and who notices when that reasoning is convincing but wrong?

The psychologist in the example is no less ethical than their colleagues. The intent is the opposite: a better assessment. The problem is that the consequences are asymmetric: the benefit is immediate and visible, the risks are delayed and diffuse. That is exactly the imbalance the study finds among physicians.

So it is not enough to say "you may not use ChatGPT in assessments". If the approved path means doing everything by hand while a tool that solves the problem is seconds away, the organisation has built an environment where the forbidden behaviour is rewarded every day.

Healthcare learned this with hand hygiene: compliance did not improve with more rules, but when the dispensers were placed where the work happens. To change behaviour, the safe option has to be the easy one.

That means healthcare needs to offer clinicians AI tools with the right security, governance and limits - and that are good enough to solve the problems that drive people to shadow AI in the first place.