Doctors, augmented: what recent studies say about artificial intelligence in medicine and a practical playbook for using tools without replacing clinicians

Executive summary

The question of whether artificial intelligence will replace doctors is one of the most pressing in modern medicine. The current evidence from 2024 and 2025 provides a clear, consistent answer: the future is one of augmentation, not substitution. AI is already delivering real gains in reducing documentation burden, accelerating knowledge retrieval, and supporting triage. However, these benefits are counter-balanced by significant, documented risks, including automation bias and the potential for clinician de-skilling.

For UK clinicians, this new era is being shaped by robust guardrails. The official NHS ambient scribe guidance and the MHRA’s AI Airlock for medical devices are creating a framework for safe, governed adoption (NHS England, GOV.UK). This article explores what the latest literature says about the AI vs doctor dynamic and provides a practical playbook for using these new tools safely and effectively.

What the latest literature actually says (2024–2025)

The academic and professional discourse has moved beyond hype to a more sober, evidence-based assessment.

Perspective & frameworks: The New England Journal of Medicine and its new sister journal, NEJM AI, caution that the perceived benefits of AI depend heavily on what it is being compared against. True value is contingent on verifiable performance, robust bias controls, and a seamless fit into existing clinical workflows (New England Journal of Medicine).
The de-skilling signal: A landmark multicentre study in The Lancet Gastroenterology & Hepatology raised a critical red flag. It reported a significant decay in the skills of endoscopists when they alternated between using AI assistance and working without it. Their adenoma detection rate dropped during their "off-AI" periods after prolonged AI use, highlighting the real-world risk of de-skilling with "always-on" helpers (Financial Times, TIME).
Clinician & public sentiment: Reporting in the BMJ and statements from bodies like the American Medical Association are converging on a common refrain: “AI won’t replace physicians—but physicians who use AI may replace those who don’t.” There is strong enthusiasm for the technology's potential, but this is heavily tempered by safety concerns (BMJ, American Medical Association).
Patient/peer perceptions: Recent studies in journals like Nature are examining how peers and patients view doctors who use generative AI. The use of AI is best accepted when it is positioned as a tool to verify a clinician's thinking, rather than as the primary source of information (Nature, Oxford Academic).
Global governance lens: The World Health Organization’s 2024 guidance on large multimodal models (LMMs) clearly outlines the key use cases but also the significant risks, including hallucinations and automation bias. The WHO calls for a global commitment to transparency and non-negotiable human oversight (World Health Organization).

Why total replacement is unlikely (near- to mid-term)

The evidence points to several fundamental reasons why AI is not poised to replace doctors in the foreseeable future.

Irreducibly human tasks: The core of clinical practice involves skills that are, for now, uniquely human. This includes nuanced shared decision-making, navigating complex ethical dilemmas, framing risk, handling safeguarding concerns, and performing a subtle physical examination.
Regulatory reality: The UK's regulatory pathways, from the MHRA AI Airlock to the NICE Evidence Standards Framework (ESF), are all built on the assumption of clinician-in-the-loop oversight and rigorous post-market monitoring, not full autonomy (GOV.UK, medregs.blog.gov.uk).
The risk surface: The dangers of automation bias (over-trusting an AI's output), the potential for adversarial inputs (deliberately tricking an AI), and the challenge of managing uncertainty all demand a final, accountable human verification step (New England Journal of Medicine).

Where artificial intelligence helps today (evidence-aligned use cases)

Ambient scribing / documentation: The NHS England 2025 guidance endorses the carefully governed deployment of ambient scribes. The benefits are clear: reduced time-to-note, improved eye contact with patients, and relief from burnout. The risks—over-trust and skill drift—must be managed by strict verification.
Knowledge retrieval with citations: Retrieval-Augmented Generation (RAG) tools can surface guideline-linked answers in seconds. These are best used for reference and education, with clinicians pasting the sources into their notes and learning logs.
Operational co-pilots: AI is proving effective at triage, classifying clinician inbox messages, preparing pre-visit summaries, and drafting discharge communications—but always with clear audit trails and mandatory approval gates (New England Journal of Medicine).

“Augment, don’t substitute”: a practical model for clinics

Default stance: The AI is a supporting tool; the clinician is the accountable decision-maker.
The verification loop: This four-step process, which mirrors the principles of the NHS ambient-scribe guidance, is essential for safe practice:
1. AI drafts: The AI generates a note, a summary, or a list of options.
2. Source check: The clinician opens the cited links to verify the information against the primary source.
3. Clinician edits/approves: The clinician makes any necessary corrections and takes final ownership of the output.
4. Log provenance: The final note should document that an AI assistant was used and cite the key sources.
Duty to maintain skills: Learning from the colonoscopy de-skilling signal, departments should consider scheduling regular "AI-off" sessions or randomised "no-assist" periods in high-acuity domains to ensure skills are maintained (Financial Times).

UK governance & assurance (what to ask vendors)

NHS England ambient scribe guidance: Any scribe tool must be a registered product with a full clinical safety case, a DPIA, and a clear human verification workflow.
MHRA AI Airlock: This supervised sandbox is the UK's pathway for testing novel AI as a Medical Device (AIaMD). Its outputs will inform future rules.
Clinician role: Your incident and near-miss reporting is now a formal part of Post-Market Surveillance. Insist on tools that provide citations, last-updated stamps, and clear override/approval logs.

How a tool like iatroX fits

iatroX is designed to be a supportive tool, not a substitute for clinical judgment.

Use it for: Getting fast, citation-first answers to your questions for reference and education. Use the Brainstorm feature to structure your thinking on differentials and investigation plans. Use the CPD feature to save your learning conversations and export PDF evidence for your portfolio.
What it is not: It is not a substitute for your own clinical reasoning or for patient-specific directives. Always pair its outputs with checks against definitive sources like NICE guidelines and your local policies.

Micro-workflows (copy-ready boxes)

Clinic unknown: Ask a cited AI tool like iatroX for a summary → open the underlying guideline links to verify → make a shared decision with the patient → document your reasoning and attach one or two key citations.
Note drafting: Use an NHS-compliant ambient scribe → verify the draft line-by-line for accuracy and omissions → add specific drug doses from the BNF → sign and file the note.
Skill-retention: Book one "AI-off" clinic or reporting session per week for high-acuity tasks. Track your performance to ensure it remains at parity with your AI-assisted work.

Risks & Mitigations

Risk Mitigation
De-skilling Rotate "AI-off" periods; targeted training on core skills.
Hallucinations Mandate citation display; require human sign-off; use tools that can "abstain" from answering.
Automation Bias Require explicit "agree/override with reason" prompts in the UI; show model confidence scores.
Data Protection Only use registered/approved tools for patient data, as per NHS guidance. Never use consumer apps.

Risk	Mitigation
De-skilling	Rotate "AI-off" periods; targeted training on core skills.
Hallucinations	Mandate citation display; require human sign-off; use tools that can "abstain" from answering.
Automation Bias	Require explicit "agree/override with reason" prompts in the UI; show model confidence scores.
Data Protection	Only use registered/approved tools for patient data, as per NHS guidance. Never use consumer apps.

FAQ

Will artificial intelligence replace doctors?
- Not based on the current evidence. The strongest signals from bodies like the WHO and from studies in the NEJM support a clinician-in-the-loop augmentation model, with robust governance.
Is there proof of harm from AI?
- Early evidence of de-skilling exists, as shown in the colonoscopy study. This is why AI-off training and mandatory human verification are so critical.
What do major UK bodies say?
- The WHO urges transparency and oversight. NHS England has set clear rules for ambient scribes. The MHRA is actively testing new AI medical devices via its AI Airlock sandbox.
How should I use iatroX?
- Use it as a powerful reference and learning tool. Leverage its citations and CPD export feature. It is not designed for autonomous, patient-specific decision-making.