Executive summary
The fear that artificial intelligence could deskill the medical profession is real and widespread. Clinicians and trainees report valid concerns about over-reliance on AI, particularly in core skills like diagnostic reasoning, and "automation bias"—the well-documented tendency for humans to over-trust automated systems—is a known risk in clinical decision support (clinicalradiologyonline.net, PMC, ScienceDirect).
However, the counter-case is equally compelling. When used correctly, AI can augment human expertise by speeding up knowledge retrieval, widening differential diagnoses, and supporting complex decisions. The evidence shows that the benefits are highly dependent on how these tools are integrated into workflows and how clinicians are trained to use them (PMC, PubMed). Crucially, robust guardrails already exist. Guidance from the World Health Organization and consensus frameworks like FUTURE-AI emphasise transparency, mandatory human oversight, and the development of new competency-based training. The path forward is not to halt innovation, but to embed AI literacy, prompt engineering, and critical appraisal skills into the UK medical curriculum, ensuring technology enhances—rather than erodes—the art of medicine.
Framing the dilemma: what clinicians mean by “deskilling”
The concern over clinical AI deskilling is not just a fear of job replacement; it's a nuanced anxiety about the erosion of core competencies. We can define it in two ways:
- Deskilling: The gradual loss of foundational clinical skills, such as pattern recognition in radiology or the ability to formulate a broad differential diagnosis from first principles.
- Upskilling inhibition: A phenomenon where trainees have fewer opportunities to acquire core skills in the first place because an AI tool consistently intermediates or automates the task (SpringerLink).
This is particularly acute in specialties like radiology, where practitioners have raised concerns about losing their expert pattern-recognition abilities if AI becomes the default first reader (clinicalradiologyonline.net, aura.abdn.ac.uk). This is compounded by the proven risk of automation bias, where consistent evidence shows that people can over-trust a decision aid, leading to errors of both omission (missing something the AI missed) and commission (accepting a flawed AI suggestion) (PMC, ScienceDirect).
Evidence check: what trials and studies actually show
The 2024–2025 literature provides a sober, mixed picture of the human vs AI doctor dynamic.
- LLMs and diagnostic reasoning: A significant randomised clinical trial published in JAMA found no significant improvement in physicians’ diagnostic reasoning when they were given an LLM assistant. In fact, the LLM alone outperformed both the junior and senior physician arms of the trial, highlighting a major gap in our understanding of how to effectively integrate these tools into the clinical workflow (JAMA Network, PubMed).
- Management reasoning: Interestingly, a companion RCT reported that LLM assistance did improve management reasoning, suggesting that different cognitive tasks (diagnosis vs. planning) benefit unequally from AI support (PMC).
- Safety-netting pitfalls: Methodological commentaries have noted instances where LLMs under-recommend seeking urgent care, underscoring the critical need for explicit clinical guardrails and human oversight in any patient-facing or triage application (JAMA Network).
The augmentation thesis: how AI can strengthen expertise
The most productive view is to frame the future of doctors with AI as a partnership. The goal is a cognitive partner, not a replacement. AI can augment clinical expertise by:
- Expanding a differential diagnosis to counter anchoring bias.
- Surfacing red-flag symptoms for a given presentation.
- Compressing the time taken to search and synthesise evidence.
The benefits arise when the clinician retains final accountability and is empowered to verify the AI's sources. Design principles that preserve skill—such as transparency of reasoning, clear estimates of uncertainty, and "provenance-first" outputs with citations—are essential for allowing clinicians to critically appraise AI recommendations. This aligns with WHO guidance and frameworks like FUTURE-AI (World Health Organization, Iris, BMJ).
Where deskilling risk is highest—and how to mitigate it
The highest risk comes from passive, uncritical use of AI. This can manifest in three patterns:
- "Silent autopilot": The unchecked acceptance of AI outputs without verification.
- "Shortcut learning": Using an AI's summary to bypass the effort of first-principles reasoning.
- "Prompt poverty": Using under-specified, lazy inputs that lead to generic and less reliable outputs.
Mitigation requires a deliberate, structured approach:
- Use structured prompts that demand ranked differentials, red flags, and citations.
- Implement tandem reading protocols, where a human-AI disagreement triggers a more detailed review.
- Schedule deliberate practice without AI—"AI-off drills"—to maintain and test baseline diagnostic skills.
Education must change: a UK-ready AI competency set
The UK Medical Schools Council, in partnership with HDR UK, has already called for data and AI skills to be integrated into the undergraduate curriculum. To prevent deskilling, this curriculum must be formalised around a core set of competencies:
- AI literacy & limits: Understanding what LLMs are (and aren't), the nature of hallucinations, and the risks of algorithmic bias.
- Prompt engineering for clinical tasks: Mastering the skill of asking structured questions to get reliable answers.
- Critical appraisal of AI outputs: The ability to assess the provenance, confidence, and applicability of an AI's suggestion.
- Safety & governance: Understanding WHO principles, UK information governance, and the importance of documenting AI use.
- Human-skills retention: Engaging in scheduled AI-off assessments and simulations to practise complex, nuanced clinical judgement.
Teaching & assessment blueprint
- Learning design: A spiral curriculum should be adopted, introducing foundational AI literacy in the pre-clinical years and moving to supervised deployment and critical appraisal in the clinical years.
- Assessment: Future OSCE and OSLER stations should include both "AI-on" (critique and adapt this AI-generated plan) and "AI-off" (demonstrate independent reasoning) modes.
- Faculty development: Medical schools and Royal Colleges must invest in upskilling educators to teach and assess these new competencies.
Service design implications for the NHS
- Procurement: Checklists should now include criteria for skill-preserving UX design, such as the mandatory display of provenance and uncertainty, and the avoidance of "one-click accept" defaults for high-stakes decisions.
- Quality metrics: Trusts should track not only productivity gains from AI but also human skill indicators, such as independent case accuracy rates, to detect deskilling early.
- CPD & revalidation: In the future, a clinician's CPD portfolio may need to include documented critical appraisals of AI outputs and periodic evidence of AI-off competence.
Risks vs Mitigations
Risk Symptom Mitigation De-skilling Drop in performance during "AI-off" periods Rotate AI-off drills; targeted training on first principles Hallucinations Confident but incorrect AI outputs Mandate citation display; require human sign-off Automation Bias Uncritical acceptance of AI suggestions Require explicit "agree/override with reason" prompts in the UI Data Protection Use of patient data in consumer apps Only use registered, NHS-compliant tools; follow IG guidance
FAQs
- Will artificial intelligence inevitably deskill clinicians?
- Not if the technology is designed and taught well. The risks, like automation bias, are real and documented, but skill-preserving workflows and new educational curricula can ensure AI enhances, rather than erodes, expertise.
- Do trials prove AI always helps?
- No, the results are mixed and highly task-dependent. The benefits are greatest when AI is used in a structured, thoughtful way with robust training.
- What belongs in the medical syllabus now?
- A core set of competencies including AI literacy, prompt engineering, critical appraisal of AI outputs, and governance—a position now endorsed by UK and international medical education bodies.
Conclusion & call to action
The deskilling dilemma is real, but it is not an inevitability. AI can either erode or enhance medical expertise depending entirely on the choices we make in how we design the tools and how we train our clinicians. The prudent and effective path forward is one of augmentation with accountability. We must actively preserve human skills, demand provenance from our technology, and train a new generation of clinicians to question the machine.
The next steps for the UK are clear. The Royal Colleges and medical schools should collaborate on a common AI competency framework and align their assessments accordingly. In parallel, NHS procurement must evolve to require skill-preserving UX and transparent, auditable outputs in all future clinical AI deployments.