AI in Medicine 2026: What Clinicians Should Know

Artificial intelligence in medicine is no longer a future topic. Regulators have authorised well over a thousand AI-enabled devices, large language models (LLMs) are being trialled for documentation and patient communication, and the EU has passed the first comprehensive AI law. At the same time, much of what is said about medical AI is marketing. This overview covers the state of play in October 2026 and what it means for practising clinicians.

1. AI-enabled devices: what the FDA list shows

The U.S. Food and Drug Administration publishes a public list of AI-enabled medical devices that have been authorised for marketing in the United States. It shows how much medical AI has actually passed regulatory review.

The numbers, with date

As of 4 September 2026, the list showed:

Entries per decision year have grown steadily. The list contains 162 devices with decisions in 2022, 226 in 2023, 235 in 2024 and 335 in 2025, plus 181 in the first half of 2026 up to the latest listed decision.

How to read the list correctly

• It is not exhaustive. The list is updated periodically; devices whose decision summaries were not published within the data-collection period are added in later updates.
• Authorisation is not proof of clinical benefit. Most entries are cleared through pathways that compare a device to an existing one. A listing tells you the device may be marketed for its stated use, not that it improves patient outcomes in your setting.
• Radiology dominates. About three in four entries sit in the radiology panel: image triage, detection support, measurement and reconstruction. If you work in imaging, you are likely already using AI, sometimes without knowing it.
• LLMs are a new frontier. The FDA plans to explore methods to identify and tag devices that incorporate foundation models, from LLMs to multimodal architectures. In other words, generative AI in regulated devices is still being worked out.

Questions to ask before using an AI device

1. What exactly is the intended use on the label, and does my use match it?
2. On which population was it validated, and does that match my patients?
3. What happens when it is wrong, and who notices?
4. How does it change my workflow: does it triage, flag or decide?
5. How are updates handled, and will I be informed when the model changes?

2. Large language models in clinical workflows

LLMs are general-purpose text models. They are not designed as medical devices, and consumer chatbots are not FDA-authorised for diagnosis or treatment. Still, they are being used and piloted for:

• Documentation: drafting notes, discharge letters and summaries, often through "ambient" tools that listen to the consultation.
• Patient communication: drafting replies to portal messages.
• Information retrieval: summarising guidelines or literature.
• Administrative work: coding suggestions, prior authorisation letters.
• Education: explaining concepts and generating practice material; see AI in medical education.

What the research shows, and what it does not

Two widely cited studies illustrate both the promise and the limits.

Patient questions. A 2023 study compared physician and ChatGPT answers to 195 patient questions posted to a public online forum. Blinded evaluators preferred the chatbot answer in 78.6% of 585 evaluations; chatbot answers were rated good or very good in quality in 78.5% vs 22.1% for physicians, and empathetic in 45.1% vs 4.6%. But the physician answers came from an online forum, were much shorter (median 52 vs 211 words), and the study did not measure patient outcomes. It shows LLMs can draft readable, empathetic text, not that they can replace clinical judgment.

Exam questions. Another 2023 study found that ChatGPT performed at or near the passing threshold on USMLE-style questions for all three Steps without specialised training. Exam questions are curated, have a single best answer and contain all necessary information. Real patients do not.

Known failure modes

• Fabrication ("hallucination"): confident, plausible but false statements, including invented references.
• Bias: outputs that reflect biases in training data.
• Automation bias: clinicians over-trusting a fluent answer.
• Privacy and security: patient data sent to external services.

The WHO names the same risks in its guidance on large multi-modal models, including false, inaccurate, biased or incomplete statements and cybersecurity vulnerabilities, and it warns about automation bias.

Practical rules for clinicians

• Use only tools approved by your employer for anything involving patient data.
• Never paste identifiable data into consumer chatbots.
• Verify every factual claim, dose and reference against a primary source.
• Document as you would without AI: you are responsible for what you sign.
• Be transparent with patients when AI is part of their communication, according to local policy.

3. Regulation

European Union: the AI Act

The EU Artificial Intelligence Act phases in its rules over several years:

The AI Act takes a risk-based approach. Under EU rules, AI that is itself a medical device or diagnostic test, or a safety component of one, counts as high-risk, so most AI medical devices fall into this category. These obligations come on top of existing medical device rules. The start dates have already been pushed back once and may change again.

What it means for clinicians: deployers of high-risk AI (which can include hospitals) will have obligations such as using systems according to their instructions, assigning human oversight to competent staff, and monitoring operation. Expect your institution to formalise AI governance.

United States

In the U.S., AI that meets the definition of a medical device is regulated by the FDA through the usual device pathways, and the public list described above is the transparency tool. Some entries are marked as including a predetermined change control plan (PCCP), which describes how a model may be modified after authorisation.

WHO guidance

On 18 January 2024 the WHO published guidance on the ethics and governance of large multi-modal models in health, with more than 40 recommendations for governments, technology companies and healthcare providers. It outlines uses in diagnosis and clinical care, patient-guided use, clerical and administrative tasks, medical and nursing education, and scientific research and drug development. Among its recommendations: governments should assign a regulator to assess and approve such models for health use, and large-scale deployments should be audited after release.

What changes for different specialties

AI adoption is uneven across medicine, and the FDA list reflects that:

• Radiology and imaging-heavy specialties see the most authorised tools: detection support, triage of urgent findings, quantification and image reconstruction.
• Cardiology has a substantial share, including ECG analysis and cardiac imaging support.
• Neurology tools include support for imaging and signal analysis.
• Primary care and internal medicine encounter AI more through documentation tools and decision support embedded in electronic health records than through stand-alone authorised devices.
• Pathology appears in the list from its earliest entries; the oldest entry on the FDA list is an automated Pap smear screening system with a 1995 decision date.

For trainees, the practical consequence is to learn how the tools in your own department work: their intended use, typical errors and how results are presented. More on diagnostic AI specifically is covered in our article on AI in diagnostics.

A checklist for departments introducing AI

• Is there a named clinical lead responsible for the tool?
• Has the tool been evaluated locally on our own patients before routine use?
• Do staff know the intended use and limitations?
• Is there a way to report errors and near misses?
• Is patient data processed in line with data-protection law and our policies?
• Will staff be told when the model is updated?
• How will performance be monitored over time?

4. Limits clinicians should keep in mind

• Validation does not travel automatically. A model that performed well in one population or scanner type can perform worse in another.
• Benchmarks are not outcomes. Exam scores and reader studies are useful, but they are not patient outcomes.
• Models change. Updates can change behaviour; ask how changes are communicated.
• Liability stays human. Regulators and professional bodies have not shifted responsibility for clinical decisions to software.
• Skills can erode. If AI handles a task routinely, keep your own skills current.

5. What this means for your career and exams

AI literacy is becoming part of professional competence, but it does not replace the knowledge you are tested on in licensing exams. You still need to recognise when an AI suggestion is wrong, and that requires the same clinical knowledge those exams assess. Evidence-based study methods such as spaced repetition and practice questions remain the most reliable way to build it. The MediFlow app uses these methods with exam-style questions written and checked against current guidelines.

Summary

• The FDA list contained 1,614 AI-enabled devices as of 4 September 2026, about three quarters in radiology.
• LLMs show promise for drafting text but carry known risks: fabrication, bias, automation bias and privacy.
• The EU AI Act classifies most AI medical devices as high-risk, with obligations for medical devices currently due from 2 August 2028.
• WHO guidance provides more than 40 recommendations for governance.
• Clinical responsibility stays with the clinician.