A family doctor finishes her last appointment at six and then opens the laptop for another hour and a half of notes, referral letters and coding. Ask clinicians what wears them down and documentation comes up as often as patient load. That's why one of the fastest-moving uses of AI in healthcare isn't a robot surgeon or a superhuman diagnostician. It's software that listens to a consultation and drafts the note.
AI in healthcare covers a wide range – from image analysis that flags a suspicious shadow on a scan to chatbots that answer appointment questions. Some of it is mature and regulated; some is experimental. This article walks through the main uses, how they're regulated, and the risks that patients, clinicians and the companies building these tools need to take seriously.
Where AI is used in healthcare today
Medical imaging and diagnosis support
Image analysis is the most established clinical use of AI in healthcare. Models trained on large sets of labelled images can highlight possible findings on X-rays, CT and MRI scans, mammograms, retinal photographs and pathology slides. Regulators such as the US Food and Drug Administration have authorised a substantial number of AI-enabled medical devices, with radiology making up the largest share.
These tools are typically decision support: they flag, prioritise or measure, and a clinician makes the call. A common pattern is triage – moving scans with suspected urgent findings, like a possible bleed, to the top of a radiologist's worklist.
Clinical documentation
Ambient AI scribes record (with consent) a consultation, transcribe it, and use a language model to draft a structured clinical note, which the clinician reviews and signs. Related tools draft discharge summaries, referral letters and replies to patient messages. The appeal is obvious: time back for clinicians, and more eye contact with patients.
Operations and administration
Less visible but widespread: predicting no-shows, optimising theatre schedules, forecasting bed demand, assisting with medical coding and billing, and handling routine patient queries.
Research and drug discovery
AI is used to analyse biological data, predict protein structures and screen candidate molecules. DeepMind's AlphaFold, which predicts protein structures, is the best-known example. Translating these advances into approved medicines takes many years and still depends on clinical trials.
How AI in healthcare is regulated
Software that is intended to diagnose, treat, monitor or prevent disease is typically regulated as a medical device.
- In the US, the FDA regulates software as a medical device and has published guidance on AI-enabled devices, including how manufacturers can plan for updates to learning algorithms.
- In the EU, such software falls under the Medical Device Regulation or the In Vitro Diagnostic Regulation. The EU AI Act additionally classifies AI that is a safety component of these regulated devices as high-risk, layering requirements on data governance, transparency and human oversight on top.
- Data protection applies everywhere: HIPAA in the US, and in Europe the GDPR treats health data as a special category requiring extra safeguards.
A tool that only drafts documentation for a clinician to review may sit in a different regulatory position than one that suggests a diagnosis – but the line depends on its intended purpose and claims, and vendors should get regulatory advice early.
The risks of AI in healthcare
Errors and hallucinations
Language models can produce fluent text that is wrong. In a clinical note, that might mean a medication that was discussed but not prescribed, a symptom the patient denied recorded as present, or a dose transcribed incorrectly. Review by the clinician isn't a formality; it's the safety mechanism.
Bias and unequal performance
A model is only as representative as its training data. If images or records came mostly from certain populations, hospitals or devices, performance can be worse for others – for example, on darker skin tones in dermatology, or on scanners from a different manufacturer. Validation should check performance across subgroups, not just overall.
Automation bias
When software is usually right, people stop checking. A clinician who has accepted fifty correct AI suggestions may wave through the fifty-first wrong one. Interface design matters here: showing confidence, highlighting the evidence and making disagreement easy all help.
Dataset shift
A model validated at one hospital can degrade at another, or at the same hospital after a change in equipment, patient mix or clinical practice. Performance needs ongoing monitoring, not a one-time sign-off.
Privacy and security
Recording consultations and sending health data to cloud services raises consent, storage and access questions. Patients should know when AI is involved, and data flows must comply with local health-privacy law.
Accountability
When an AI-assisted decision goes wrong, responsibility can be unclear between the clinician, the hospital and the vendor. Clear policies on who reviews what, and documentation of that review, protect everyone.
A checklist for adopting AI in a clinical setting
| Question | Why it matters |
|---|---|
| Is it cleared or certified for this intended use? | Regulatory status depends on the specific claim |
| Was it validated on a population like ours? | Guards against bias and dataset shift |
| Who reviews the output before it affects care? | Human oversight is the main safety net |
| How are errors reported and tracked? | You need to know when it fails |
| Where does patient data go, and for how long? | Privacy law and patient trust |
| Do patients know AI is involved? | Consent and transparency |
| What happens if the vendor's model is updated? | Changes can alter performance |
The bottom line
AI in healthcare is delivering real value today in narrow, supervised roles – flagging findings on images, drafting documentation and smoothing administration. The risks are equally real: confident errors, uneven performance across patient groups, over-reliance and privacy exposure. The organisations getting this right treat AI as a tool that supports clinicians rather than replaces their judgement, validate it locally, monitor it continuously and keep patients informed.
