Medicine is where AI's promise and its difficulty are both at their sharpest. The data is rich, the stakes are high, and the gap between "works on a benchmark" and "helps a patient" is wider than in any other field. After a decade of headlines, a clearer picture has emerged of where AI has genuinely cleared the bar, where it's still a demo, and what determines the difference.
Where AI has actually arrived
- Medical imaging. This is the mature category. The US FDA has authorised well over a thousand AI-enabled devices, the large majority in radiology: flagging intracranial haemorrhage or pulmonary embolism on CT for prioritisation, detecting diabetic retinopathy from retinal photographs (the first autonomous diagnostic AI, cleared in 2018), assisting mammography reads, and measuring anatomy that previously took manual effort. Large trials in breast screening have shown AI-supported reading matching or exceeding double human reading with less radiologist time.
- Clinical documentation. Ambient scribes — systems that listen to the consultation and draft the note — went from pilot to widespread deployment in 2024–25. They address the single biggest source of clinician burnout, and their errors are visible to the clinician before signing, which is a favourable safety profile.
- Triage and deterioration prediction. Models that watch vital signs and labs to flag sepsis or clinical decline earlier than staff would. Results are mixed and depend heavily on implementation; some widely deployed sepsis models were found to perform far worse in practice than their vendors claimed.
- Drug discovery. Protein-structure prediction (AlphaFold and successors) changed structural biology as a field. AI-designed molecules have reached clinical trials; none is yet an approved drug, but the pipeline is real.
- Operations. Scheduling, bed management, coding and billing — unglamorous, measurable, and often the fastest return.
Why the demo-to-deployment gap is so wide
A model that scores brilliantly on a curated dataset frequently fails in a real hospital, for reasons that are now well understood.
- Distribution shift. Different scanners, patient populations, protocols and prevalence. A model trained at one academic centre can lose a large fraction of its accuracy at a community hospital.
- Label leakage and shortcuts. Models learn whatever predicts the label. Famously, a pneumonia classifier learned to recognise the portable X-ray machine used for the sickest patients rather than the pneumonia.
- Bias baked into the data. Historical care is unequal, so models trained on it inherit the inequality. A widely used risk score that used healthcare cost as a proxy for need systematically under-referred Black patients because less had been spent on their care.
- Workflow fit. An alert nobody has time to act on is noise. Alert fatigue kills more clinical AI projects than accuracy does.
- Automation bias. Clinicians defer to the machine, including when it's wrong. The system's net effect depends on how the human-AI pair performs, not on the model alone.
Regulation and validation
Diagnostic and treatment AI is regulated as a medical device in most jurisdictions — by the FDA in the US, under the MDR and the AI Act's high-risk provisions in the EU. Regulators have adapted: the FDA's framework for predetermined change control plans lets a model be updated within pre-agreed bounds without a new submission, acknowledging that models evolve. What good validation looks like is also converging: external validation on data from sites not used in training, prospective evaluation in the real workflow, performance reported by subgroup, and post-market monitoring for drift. Reporting standards such as TRIPOD-AI and CONSORT-AI exist so that studies can be compared.
Large language models in the clinic
General-purpose LLMs pass medical licensing exams and draft patient letters convincingly, which has created both enthusiasm and a specific hazard: fluent, confident, occasionally wrong output in a setting where wrong is dangerous. The deployments that have gone well share a pattern — the model drafts, retrieves or summarises, and a clinician remains the decision-maker with the source material one click away. Direct patient-facing medical advice from general models remains contentious, and most health systems restrict it. Privacy is the other constraint: patient data entering a third-party model needs a contractual and technical basis, which is why on-premises and private deployments dominate.
What determines whether it helps
The projects that deliver measurable benefit tend to have: a clearly defined clinical question with a clear action when the model fires; validation on local data before go-live; a workflow designed with the people who'll use it; monitoring of both model performance and clinical outcomes after launch; and governance that can switch the system off. The technology is rarely the bottleneck anymore. The organisation is.
- Imaging, documentation and operations are where AI has proven value; prediction models are hit-and-miss in practice.
- Distribution shift, shortcut learning and data bias make external, prospective validation essential.
- Clinical AI is regulated as a medical device; frameworks now allow controlled model updates.
- Human-in-the-loop designs where the clinician decides and can see the evidence are the pattern that works with LLMs.