Where AI is actually deployed today
The largest cluster by far is medical imaging. Algorithms triage scans so that a suspected large-vessel stroke or intracranial bleed rises to the top of a radiologist's list, detect nodules on chest imaging, flag possible fractures, and support mammography reading. These systems generally act as a second pair of eyes or a queue-sorter rather than a decision-maker: the radiologist still reports the study, and the software's job is to make sure the urgent case is not sitting eightieth in the queue.
The second cluster is ambient clinical documentation. A microphone captures the consultation and a language model drafts the note, which the clinician edits and signs. This has spread quickly because the problem it solves is enormous and unglamorous — clinicians spend a substantial share of the working day on documentation — and because an error in a draft note is visible to the person who has to sign it. The third is deterioration and sepsis prediction in hospitals, where a model watching vital signs, labs and orders raises an alert hours before the pattern is obvious. Fourth is retinal screening for diabetic retinopathy, one of the first areas where an autonomous system was authorised to produce a screening result without a specialist reading every image.
The part you will never see: discovery
The most consequential AI in medicine may be the part furthest from the clinic. Machine-learning systems for predicting protein structure changed structural biology within a few years, turning a problem that could take a laboratory months into something that runs computationally, and the predicted structures are now a routine starting point in academic and commercial drug discovery. Related models are used to search chemical space for candidate molecules, predict binding, and prioritise which compounds are worth synthesising at all.
The honest caveat is that this compresses the earliest and cheapest part of drug development. A candidate still has to survive animal work, phase 1 safety, phase 2 signal-finding and phase 3 outcome trials, and the overwhelming majority still fail there for reasons no model currently predicts — toxicity in humans, or simply not working. Several AI-originated molecules have entered clinical trials, which is a genuine milestone, but a molecule in trials is not a medicine. Expect the effect on approved drugs to arrive slowly, and to be visible first as more shots on goal rather than a higher scoring rate.
What regulation actually looks like
In the United States, clinical AI is regulated as a medical device. Most products reach the market through the 510(k) pathway by demonstrating substantial equivalence to an existing device, with a smaller number going through De Novo or premarket approval. The FDA maintains a public list of AI-enabled and machine-learning-enabled devices it has authorised; that list now runs into the many hundreds and is dominated by radiology, with cardiology a distant second. Authorisation means the device met a regulatory standard for its stated purpose — not that it improves outcomes in your hospital, on your population, in your workflow.
In the European Union, medical AI sits under both medical device regulation and the AI Act, which classifies most health applications as high risk and attaches obligations covering risk management, data governance, technical documentation, human oversight, accuracy and post-market monitoring. Separately, standards bodies have been building the measurement scaffolding — NIST's AI Risk Management Framework is the most cited voluntary reference in the US. A live regulatory question everywhere is how to handle models that continue to learn after approval, since the traditional device framework assumes the thing you approved stays the thing you shipped.
The real failure modes
Dataset shift is the most common. A model trained on one hospital's scanners, patient mix and labelling conventions quietly degrades when deployed somewhere else, or in the same place after a scanner is replaced or a coding practice changes. Performance decays without anyone noticing, because nothing announces it. This is why serious deployments include ongoing monitoring rather than a one-off validation, and why a published accuracy figure from the developer's own data tells you less than it appears to.
Automation bias is the second: people defer to a confident machine, including when it is wrong, and the effect is strongest for less experienced users and under time pressure. A tool that improves average performance can still cause specific harms by anchoring a clinician away from the right answer. Third, models trained on non-representative populations perform worse for the groups that were underrepresented, which in practice compounds existing inequities. Fourth, large language models hallucinate — producing fluent, confident, entirely fabricated content, including citations. In a drafting tool with a human editor that is a nuisance; in a system giving advice directly to patients it is a hazard, which is precisely why general-purpose chatbots are not authorised as diagnostic devices.
What this means for you today
Practically: if you have had imaging done in the last few years, software may well have been involved in prioritising or double-reading it, and a human still made the call. If your clinician used a laptop or phone to record the consultation, an ambient documentation tool may have drafted the note — you can ask, and you can decline. If you were admitted to hospital, a background model may have been watching your observations. None of these replace clinical judgement, and none of them are things you need to manage.
What you should not do is treat a general-purpose chatbot as a clinician. These systems are genuinely useful for preparing questions, understanding terminology and making sense of a diagnosis you already have. They are not evaluated as diagnostic devices, they cannot examine you, they will produce a confident answer regardless of whether they have grounds for one, and they are systematically bad at knowing when a symptom needs urgent attention. If you want help preparing for an appointment, our appointment prep tool and the questions to ask resource are built for exactly that.
Questions worth asking about any medical AI
Four questions separate a serious product from a demo. What exactly is the task, stated narrowly? Who was it validated on, and does that population resemble the people it will be used on? Is a human reviewing the output, and does that human have the information and the time to disagree? And what happens when it is wrong — is there monitoring that would detect a slow decline in performance, and who is accountable for the result?
A fifth question matters for anything consumer-facing: what happens to the data. Health information you type into an app is often not covered by clinical privacy law, and the terms you agreed to may permit uses you would not expect. Our digital health hub covers that in detail. If a product cannot answer the first four questions clearly, the safest assumption is that it has not been asked them.