The Deskilling Question · Essay 1

Will AI make doctors worse at thinking?

AI can improve a doctor’s performance today without proving that it preserves the doctor’s ability to perform tomorrow. Clinical AI should therefore be judged not only by the answers it improves, but by the judgment it leaves intact.

The wrong question is whether AI improves accuracy

Editorial position: immediate task performance and retained independent competence are different outcomes.

The familiar evaluation is simple: give clinicians cases, give half of them an AI tool, and compare the final scores. If the assisted group performs better, the tool appears useful. That is an important question—but it is not the deskilling question.

A doctor can reach a better answer with AI today while becoming less able to reach that answer alone next year. A system can reduce workload while also reducing exposure to the difficult cases through which expertise is formed. It can make care safer at one point in the workflow yet make an undetected error more likely when the system fails. None of these possibilities is resolved by a single-session accuracy comparison.

The useful question is therefore broader: what happens to human judgment before, during, and after repeated use? Does AI help the clinician notice more, reason better, and learn—or does it encourage the clinician to accept, verify superficially, and gradually stop practising the very skill needed to recognize a wrong answer?

Automation bias, cognitive offloading, and deskilling are not synonyms

Evidence boundary: the mechanisms overlap, but they require different study designs and should not be reported as interchangeable effects.

Automation bias is a decision error in the presence of an automated recommendation: accepting an incorrect suggestion, or failing to act because the system did not issue an alert. It can occur immediately, even in an expert who has lost no underlying skill.

Cognitive offloading is the transfer of work from memory, calculation, search, or interpretation to an external tool. Offloading is not inherently harmful. Surgeons already use checklists, navigation, references, and calculators. The relevant question is whether the work being offloaded is routine and recoverable—or whether it is the repeated practice through which clinical judgment is maintained.

Deskilling is a decline in the ability to perform a task independently after repeated reliance on automation. Proving it requires time, prior exposure measurement, and an unassisted test. A cross-sectional experiment showing that a doctor followed bad AI demonstrates automation bias; it does not by itself demonstrate that the doctor’s skill has atrophied.

This distinction protects the debate from two symmetrical errors. We should not call every AI-related mistake “deskilling.” We should also not infer that skill is preserved merely because assisted performance remains high.

LLMs can help—but collaboration is contextual, not automatic

Evidence: randomized and controlled vignette studies; useful for task performance, not long-term skill retention or patient outcomes.

A 9 September 2026 npj Digital Medicine randomized field-laboratory study gave 249 physicians in Indonesia, Kenya, and the Netherlands access to GPT-4o while they worked through clinical cases. Access was associated with gains of 10.7 percentage points in Indonesia, 18.0 in Kenya, and 7.2 in the Netherlands. The context matters: these were controlled cases, the control groups did not have internet or guideline access, patient harms were not assessed, and the study did not test retained unassisted skill. Read the study.

Other trials complicate the story. In a 2024 JAMA Network Open randomized trial of 50 physicians, diagnostic reasoning scores were 76% with an LLM and 74% with conventional resources; the adjusted difference was 2 percentage points (95% CI, −4 to 8; P=.60). The LLM alone scored 16 points higher than physicians using conventional resources. Giving doctors access to a capable model did not automatically produce the best human–AI team. Read the trial.

Training appears to matter. A 2026 randomized trial in Pakistan included 58 physicians who first completed 20 hours of AI-literacy training. The assisted group scored 71.4% versus 42.6% without LLM access, an adjusted difference of 27.5 points. Yet the LLM alone still outperformed assisted physicians by 11.5 points, and the cases remained simulated rather than clinical outcomes. Read the trial.

Taken together, these studies show that AI can improve case performance, but the size of the benefit depends on setting, comparison resources, training, task design, and how clinicians use the output. They do not establish that repeated use preserves diagnostic reasoning, examination skills, or independent decision-making over time.

The dangerous experiment begins when the AI is wrong

Evidence: controlled reader and learner experiments; demonstrates susceptibility to misleading advice, not longitudinal skill loss.

Correct AI and incorrect AI are not mirror images. Helpful suggestions may produce modest gains because clinicians already recognize many correct answers. A confident wrong suggestion can instead pull attention toward the wrong feature, narrow the differential, and make a plausible error easier to accept.

In a Radiology study of 27 radiologists reading 50 mammograms, deliberately incorrect AI suggestions impaired readers at every experience level. Accuracy under erroneous suggestions was approximately 19.8% among inexperienced readers, 24.8% among moderately experienced readers, and 45.5% among very experienced readers, compared with roughly 80% when the AI advice was correct. Greater experience was protective, but not protective enough. Read the study.

A 2026 npj Digital Medicine experiment in 111 novice medical students found that misleading AI explanations significantly reduced accuracy, while correct explanations did not significantly improve it over no explanation. Confidence did not reliably separate correct from incorrect responses. Read the study.

Explanation alone is therefore not a safety feature. A polished rationale can help a clinician interrogate a recommendation, but it can also increase the persuasive force of a wrong one. The relevant endpoint is not whether the AI produced an explanation; it is whether the human detected and corrected the error.

The evidence for actual clinical deskilling is still thin

Evidence: one important retrospective real-world signal; causal interpretation remains uncertain.

Longitudinal clinical evidence is much rarer because it requires a difficult design: measure clinicians before automation, observe sustained exposure, then test the same skill without assistance while accounting for case mix, training, staffing, and secular change.

The strongest current warning signal comes from colonoscopy rather than AI neurosurgery. A 2025 multicentre observational study examined 1,443 colonoscopies performed without AI at four Polish centres before and after clinicians had routine exposure to AI-assisted colonoscopy. Adenoma detection rate during the non-AI procedures declined from 28.4% to 22.4%, an absolute difference of −6.0 percentage points (95% CI, −10.5 to −1.6; P=.0089). Read the corrected publication record.

That is a real-world association and should not be dismissed. It is also a retrospective observational before-and-after study, not causal proof that AI exposure produced skill decay; the design cannot establish causation. Unmeasured changes in patients, workflow, or clinicians may contribute. The responsible conclusion is narrow: independent performance can worsen after routine AI adoption, and prospective monitoring should be designed before deployment rather than added after a concern appears.

The verification paradox

Synthesis: grounded in automation-bias literature and the emerging clinical evidence above.

The standard reassurance is that AI will draft and the doctor will verify. But verification is not a lightweight residual task. To verify a generated diagnosis, segmentation, operative plan, or consent explanation, the doctor must retain much of the knowledge and attention needed to do the task in the first place.

This creates a paradox: the skill required to catch an AI error is the skill most likely to weaken when the AI performs that work routinely. Rare errors make the problem harder. When the system is correct 98 times, a clinician may scan rather than reconstruct the reasoning on the ninety-ninth. Yet the hundredth case may fail precisely because it contains an anatomical variant, contradictory record, or atypical disease presentation that the system has not learned to handle.

The answer is not to force doctors to repeat every task manually. It is to identify which cognitive acts constitute clinical responsibility and preserve them deliberately: problem representation, direct review of original data, detection of discordance, causal plausibility, exception handling, and the final decision under uncertainty.

Design for skill preservation, not perpetual dependence

Design proposals: not yet validated as a complete intervention; each requires prospective evaluation.

1. Ask for an independent assessment before revealing the AI output. In high-stakes interpretation, record the clinician’s problem representation or provisional decision first. This makes disagreement visible and reduces anchoring.
2. Measure assisted and unassisted performance separately. A deployment can improve throughput while independent competence declines. Both outcomes belong on the safety dashboard.
3. Preserve direct access to original evidence. Images, waveforms, examination findings, source passages, and model inputs should remain one step away—not hidden behind a generated summary.
4. Record change, correction, and override. Did AI change the clinician’s decision? Was that change correct? Could the clinician identify a deliberately seeded or naturally occurring error?
5. Schedule AI-off cases and drills. Periodic unassisted cases can reveal dependency before a real outage or edge case does. For trainees, protected independent practice is part of education, not inefficiency.
6. Use automation first where failure is reversible and auditable. Retrieval, structured drafting, completeness checks, and administrative transformation are safer starting points than autonomous diagnosis or operative recommendation.
7. Make uncertainty and disagreement productive. The system should show why it is uncertain, surface missing data, and invite reconciliation—not smooth every conflict into one fluent answer.

What this means for Neurosurgery.ai

A consent agent should not replace the surgeon’s formulation of the case. It can retrieve approved explanations, check whether required risks were discussed, adapt language after the clinical content is fixed, and test patient understanding. It should record what the surgeon corrected and defer when the operative plan or evidence is unresolved.

Imaging support should keep the original study visible and make comparison with the AI interpretation easy. A risk engine should calculate from a locked, validated model; the language model may explain the output but must not invent a more precise number. The observability layer should watch not only model errors, but changes in human behaviour: falling override rates, superficial review, delayed recognition, and declining unassisted performance.

In neurosurgery, the decisive variable often lies outside the average case: the poor dissection plane, variant vessel, discordant tract, change in monitoring, or patient value that alters the acceptable trade-off. AI is most useful when it expands what the clinician can see and test. It becomes dangerous when fluency makes the search for that decisive variable feel unnecessary.

Bottom line

The evidence does not justify the claim that AI is already making doctors broadly less capable. It does justify a more demanding standard than assisted accuracy. Wrong AI can redirect correct human judgment; access alone does not guarantee effective collaboration; and early real-world evidence suggests that independent performance after routine exposure deserves direct study.

Use AI to extend judgment—not to surrender it.

The best clinical AI will not be the system that makes thinking unnecessary. It will be the system that helps clinicians think at a higher level while preserving the ability to recognize when the system is wrong.

Editorial note: evidence was reviewed through 13 September 2026. The Brief distinguishes short-term assisted performance, automation bias, cognitive offloading, and longitudinal deskilling. Controlled vignette studies do not establish patient benefit or retained clinical competence; the colonoscopy finding is observational and does not establish causation. This Brief is professional information only and is not clinical advice.