AI may change the skills clinicians practise and retain, but the evidence does not justify saying that all medical AI inevitably deskills doctors. Some research raises credible concerns about performance without assistance; other studies show improved performance with assistance. Those findings address different questions and should not be forced into a single verdict.
The central distinction is between doing better while a tool is available and maintaining the ability needed when it is unavailable, wrong or outside its scope. A useful clinical system should be assessed on both, with methods appropriate to the skill being discussed.
What would count as deskilling?
Deskilling means a reduction in an established capability, not simply choosing an efficient way to complete a task. A doctor does not demonstrate lost competence merely by using a calculator or consulting guidance. The important question is whether a necessary ability has deteriorated and whether that matters for safe practice.
Failure to acquire a skill is different again. A trainee who repeatedly receives completed interpretations may have fewer opportunities to develop independent reasoning, even if no previously established expertise has been lost. That is a training-design hypothesis requiring direct assessment.
Temporary unfamiliarity also deserves separation. A clinician may remain knowledgeable but take longer to navigate a rarely used manual workflow. Conversely, someone may operate the interface confidently while failing to notice an inappropriate recommendation. Neither login activity nor perceived confidence establishes the underlying competence.
Before discussing evidence, specify the capability: visual search, history-taking, interpretation, management planning, communication or recovery from interrupted automation. The word "thinking" is too broad to serve as a measurable outcome.
The colonoscopy study is a warning signal, not a universal result
A multicentre observational study published online on 12 August 2025 examined non-AI-assisted colonoscopy before and after the introduction of AI support. The reported adenoma detection rate fell from 28.4% before exposure to 22.4% afterwards, an absolute difference of 6 percentage points. There were 795 examinations in the before period and 648 afterwards. See the original study record.
This was a before-and-after observational comparison of a specific procedural task. It raises a serious question about unaided performance after exposure, but does not establish that conversational AI causes permanent deterioration in general clinical reasoning. Nor does it establish that the observed change must be explained by skill loss rather than other differences between periods or participants.
The distinction matters. A study can identify a clinically relevant signal without resolving every causal explanation. Dismissing it because it is not definitive would be unhelpful; generalising it to every AI product and every doctor would be equally unwarranted.
Better assisted performance is a different outcome
A randomised trial published in Nature Medicine on 5 February 2025 studied 92 physicians undertaking simulated management-reasoning tasks. GPT-4 assistance improved scores compared with conventional resources in that setting. It did not establish long-term retained competence or patient benefit in routine care. See the published manuscript.
These findings can coexist with concern about deskilling. An aid may improve today's output while changing tomorrow's opportunities to practise. Alternatively, it may expose clinicians to useful alternatives and improve subsequent understanding. The direction cannot be settled merely by observing that a tool was used.
The educational question is what the user does during the interaction. Does the clinician formulate a view, compare alternatives and examine disagreement, or accept a completed answer without engaging with its basis? These are proposed mechanisms to investigate, not claims that one interaction pattern has already been shown to protect every clinical skill.
Why a polished review can conceal limited understanding
Consider a fictional training exercise. A junior doctor receives a well-written summary and a plausible proposed plan. They identify a spelling error, approve the document and describe the experience as helpful. The final output looks competent.
A supervisor then asks what missing information would change the plan. The learner cannot explain it. That does not prove the AI caused the gap: the knowledge may never have been secure. It does show why output quality alone is an incomplete measure of learning.
Now consider a different learner using the same material. They explain their initial view, identify an unsupported assumption, check the relevant source and revise the plan. The final document may look similar, but the educational activity is different.
A training programme should therefore assess reasoning directly rather than infer it from the quality of assisted documentation. The goal is not to make learners reproduce every administrative task manually. It is to preserve the tasks through which necessary clinical understanding becomes visible.
What a convincing longitudinal study would require
A proposed study would define the target skill and measure it before substantial exposure. It would compare groups with sufficiently similar experience, case mix and training opportunities, while recording how assistance was actually used.
Assessment should include delayed performance on new cases, not simply immediate repetition of a familiar exercise. It should distinguish correct answers, explanations, recognition of uncertainty and the ability to recover when a suggestion is wrong. Confidence can be recorded separately rather than used as a substitute for performance.
The comparator should retain ordinary professional resources. Removing guidance, calculators and supervision from one group would create an artificial contest unless that restriction was itself the research question.
Where unaided testing is appropriate, educational or simulated settings can protect patients from unnecessary risk. Clinically necessary support should not be withheld from real care merely to produce a cleaner experiment. Study design needs the relevant ethical and clinical oversight.
Preserve capability proportionately, not nostalgically
A practical response is to identify the abilities clinicians must retain in their actual roles. Someone supervising generated recommendations needs to recognise missing information, judge applicability, challenge unsupported certainty and know when further assessment is necessary.
Protected unaided exercises can be brief and targeted. A learner might first explain a fictional case, then inspect an AI alternative, justify any change and revisit a different case later without the earlier answer visible. A supervisor can examine the reasoning rather than count completed sessions.
Fallback practice serves another purpose. A team can rehearse continuing work when an automated service stops, including finding the current plan and distinguishing completed from pending actions. This tests recovery capability, not only knowledge recall.
These are proposed educational and operational safeguards to evaluate. They should not be marketed as a proven antidote to deskilling simply because they sound plausible.
Where iatroX fits, and where the claim must stop
As described on 10 October 2026, iatroX's Socratic Tutor opens around an attempted question and uses follow-up questions to explore the learner's misconception. The question-bank and simulation formats provide additional settings for deliberate practice.
That design offers a relevant way to organise active learning, but it does not establish that iatroX prevents clinical deskilling. Improving a platform score is not equivalent to demonstrating maintained competence in a clinical role, and simulation cannot replace supervised bedside or procedural experience.
A useful learning record would state what the learner initially thought, what changed and what they still need to practise. It should not claim a competence gain solely because an AI conversation occurred.
The cognitive cost of an AI-native workflow
The question becomes more specific when AI selects the history, constructs the summary, retrieves evidence and prepares the next action before the clinician sees the case. The possible cost is not merely reduced recall. It is reduced opportunity to decide which information matters and to notice that the framing itself may be incomplete. This is a hypothesis about workflow design, not an observed effect of Doctolib, OpenEvidence or iatroX.
In a fictional teaching exercise, the assistant presents a coherent account of recurring symptoms. Every subsequent suggestion follows logically from that account, but the summary omitted a change in timing that could alter the interpretation. A learner who reviews only the final recommendation may never encounter the point at which the case was narrowed.
A more informative exercise would allow inspection of the original history and ask the learner which facts should have remained prominent. The assessment would concern information selection and clinical interpretation, not whether the learner can produce an equally polished paragraph. The practical aim is to preserve the thinking necessary to supervise assistance without making clinicians repeat every mechanical task.
Beneficial offloading and educational loss can coexist
Removing repetitive typing or locating a familiar document could release attention for listening, examination and shared decisions. That is a plausible benefit of better software. It should not be dismissed merely because some cognitive activity has been delegated.
However, the use of released time matters. A training environment could replace routine drafting with deliberate discussion of the case, or fill the same period with additional activity that offers little feedback. The educational effect would depend partly on the surrounding organisation rather than the model alone.
A proposed evaluation should therefore record exposure to clinical reasoning opportunities as well as product use. Relevant questions include whether trainees formulate initial interpretations, receive feedback on disagreements and practise unfamiliar cases. Counting AI sessions, completed notes or time saved would not establish what was learned.
The Johns Hopkins collaboration is a reason to study the question
On 13 October 2026, OpenEvidence and Johns Hopkins Medicine announced a collaboration spanning clinical care and medical education. The announcement describes educational ambitions while retaining the importance of clinical foundations. It is not evidence that AI has already caused deskilling, nor that a completed new curriculum has solved the problem. Johns Hopkins Medicine's announcement.
The useful response is to make the learning objectives explicit. A learner should be able to retrieve relevant evidence, explain why it applies, identify uncertainty and communicate a defensible plan. Assistance can be included, but performance with assistance and retained capability should remain separate outcomes.
A proposed practice cycle for preserving independent judgement
An educational sequence could begin with an independent attempt, continue through comparison with an assisted interpretation and end with explanation of what changed. A later, different case would test whether the learner retained the distinction rather than remembered the previous answer.
As described on 15 October 2026, iatroX's question banks, Tutor and simulations provide formats that can be used within such a sequence. The Tutor focuses on follow-up questions around an attempted answer; simulation feedback can be linked to the learner's transcript. Those product designs should not be treated as direct evidence of reduced deskilling or improved patient outcomes.
Learning records should describe genuine activity and the learner's own interpretation. A reviewed CPD record can document what was practised and what still needs work, but it does not establish accredited CME or competence to supervise every AI agent. The standard remains the same: assess the skill that the clinical role actually requires, and avoid substituting polished output for demonstrated understanding.
The decision for clinical teams
Do not treat every reduction in manual effort as educational harm. Some work can be simplified without sacrificing the judgement needed to care for patients. Equally, do not assume that retaining a clinician's name beside an output means the underlying capability is being maintained.
The sensible standard is specific: demonstrate the assisted benefit, identify the competence needed to supervise the system and assess whether that competence remains. That is more demanding than either celebrating automation uncritically or defending every existing task as indispensable training.
Frequently asked questions
Has AI been proved to deskill all doctors?
No: the evidence concerns particular tasks and settings, and does not establish a universal or irreversible effect. The colonoscopy study is an important observational warning signal, not proof about every clinical AI system.
Can AI improve performance and still create a training concern?
Yes: immediate assisted performance and retained independent ability are different outcomes. Both need assessment when a task is important for future supervision or clinical practice.
Does using an AI tutor prevent deskilling?
That would require direct evidence for the relevant learners, skills and follow-up period. Reasoning-focused practice is an educational design choice, not a guaranteed protective effect.