Heidi II could change which clinical skills doctors practise, but there is no basis in the public evidence reviewed on 29 September 2026 to conclude that it causes deskilling. Evidence from other systems raises legitimate questions. It does not establish that Heidi II erodes clinical reasoning or that every reduction in unaided performance is unacceptable.
The distinction matters because immediate assistance, reliance on assistance and deterioration of independent skill are different outcomes. Treating them as interchangeable makes it harder to recognise both useful automation and genuine problems.
Define deskilling before looking for it
For this discussion, deskilling means deterioration of an established capability over time. A doctor who previously recognised a particular kind of inconsistency but can no longer do so would fit that concern. A trainee who never acquired the capability presents a different educational problem: missed learning rather than loss of expertise.
Dependence is also not automatically deskilling. A clinician can use a reference resource to achieve a result they could not produce from memory, without having lost any prior skill. The relevant question is whether they retain the ability required to judge applicability, recognise uncertainty and recover when support is unavailable or wrong.
Software unfamiliarity is another separate issue. Difficulty finding a task after an interface change does not necessarily show deterioration in clinical reasoning. An evaluation that merges these outcomes could report apparent harm while measuring the wrong thing.
The colonoscopy finding deserves attention and restraint
A multicentre observational study published in 2025 reported that unaided adenoma detection fell from 28.4% before exposure to AI-assisted colonoscopy to 22.4% afterwards. These were adenoma detection rates in the study's non-AI-assisted examinations, not percentages of clinicians who became deskilled. The source is Budzyn and colleagues' original study.
The result is a meaningful clinical warning signal. However, an observational before-and-after comparison does not by itself establish why performance changed. It concerned endoscopic visual detection, not language-model agents or general consultation reasoning. It also does not demonstrate permanent deterioration across medical practice.
For Heidi II, the appropriate inference is methodological: retained performance should be examined rather than assumed. It is not legitimate to transplant the reported detection rates into a forecast about doctors using an unrelated product.
Which capabilities might become less rehearsed?
Heidi's 28 September 2026 announcement describes agents handling work around the visit, including preparation and follow-up. The new capabilities are not available in the UK or EU. Their cognitive consequences will depend partly on which activities users delegate and how they supervise the result.
One hypothesis is reduced rehearsal of case synthesis. A clinician repeatedly receiving a prepared account might spend less time deciding which details belong together. Another is reduced practice in reconciling contradictions if the interface resolves them before the clinician sees them. A third concerns maintaining an account of unresolved work when the system manages it in the background.
These are candidate mechanisms, not findings about Heidi II. Automation might instead make important contradictions more visible, improve follow-through and release attention for interpretation. An evaluation should be capable of detecting that beneficial pattern too.
A fictional case that distinguishes the outcomes
Imagine an experienced doctor who uses an agent to organise correspondence before a clinic. With the tool available, the doctor identifies outstanding issues promptly and completes the session with less unfinished work. That would be an immediate assisted benefit.
In a later simulation without that agent, the doctor takes longer to locate the relevant information but reaches a sound interpretation using normal reference materials. This might reflect the loss of a useful organisational aid, not impaired clinical competence.
Now consider a different result: despite having the same information and adequate time, the doctor repeatedly fails to notice that an earlier plan has been superseded. If this represents deterioration from their own baseline and exceeds changes in a suitable comparison group, it would be more concerning.
The distinction turns on what changed, not simply whether the doctor performed better with the software present. It also separates this question from iatroX's discussion of the out-of-the-loop clinician, which concerns missing immediate workflow context rather than deterioration of established skill.
What a longitudinal evaluation should measure
We would begin with baseline assessments matched to the proposed delegated work. These should include synthesis, contradiction detection, explanation of alternatives and recovery from an interrupted task. Participants would then be followed during actual, appropriately governed use, with delayed simulated assessments using comparable but unfamiliar cases.
A comparison group should retain ordinary clinical support. Removing normal reference materials would create an artificial test of memory rather than a realistic assessment of professional competence. Evaluators should also record experience, workload, case complexity and how much each participant actually used the relevant feature.
The outcome should not collapse all skills into one score. A clinician might improve at evaluating evidence while becoming less familiar with the mechanics of a previously manual process. Whether that trade-off is acceptable depends on the ability needed for safe supervision and fallback.
No such Heidi II longitudinal study has been run for this article. Any claimed results should come from an actual documented evaluation, not from these proposed methods.
Retention does not mean preserving every manual task
A policy requiring doctors to continue every repetitive activity merely to preserve familiarity would be difficult to justify without showing its educational value. Re-entering identical information and interpreting an ambiguous patient account are not the same kind of work.
A more useful approach would identify capabilities that must remain dependable even when software usually assists. These may include recognising a changed situation, understanding why a plan was agreed, detecting an unsupported inference and taking over unfinished work. Practice could then be protected around those capabilities rather than around every historical administrative routine.
Supervisors should also distinguish experts from learners. An established clinician may need occasional recovery exercises; a trainee may need more opportunities to build an initial interpretation. Neither group should be assumed immune to error or incapable of benefiting from assistance.
What clinicians should conclude now
The current evidence supports vigilance, not a declaration that Heidi II deskills doctors. The product-specific question remains open because studies of different systems and tasks cannot answer it directly. At the same time, a supplier's assurance that clinicians remain responsible does not remove the need to study what competence they retain.
For clinical leaders, the priority is to specify the skills required to supervise the chosen workflow and measure them over time. For educators, it is to preserve meaningful opportunities to interpret cases. For users, it is to distinguish "the tool helped me complete this" from "I can recognise when its output does not fit".
That approach allows automation to remove unnecessary work without treating clinical understanding as an unnecessary cost.
Frequently asked questions
Has Heidi II been shown to deskill doctors?
No longitudinal Heidi II evidence establishing deskilling was identified in the public materials reviewed on 29 September 2026. Research involving other systems raises relevant questions but does not establish this product-specific outcome.
Does performing worse without AI prove deskilling?
No, it may simply show that the AI was useful. Demonstrating skill deterioration requires an appropriate baseline, comparable assessments and consideration of alternative explanations.
Can clinical practice exercises guarantee protection against deskilling?
No, educational exercises should be evaluated rather than advertised as a guaranteed antidote. Their practical purpose is to maintain and assess the understanding needed to challenge and recover from assisted work.