AI in the NHS: what is actually deployed, and what is hype

Two versions of the NHS are described in public. In one, artificial intelligence already reads the scans, drafts the notes and runs the waiting list. In the other — the one most staff actually work in — the software on the ward is a slow electronic record, and the closest thing to AI is a triage tool in radiology that some consultants trust and some quietly ignore. Both descriptions contain truth. Telling them apart is the whole skill this topic tests.

Panels ask about AI constantly, and they hear the same answer back: it will not replace doctors, but doctors who use AI will replace those who do not. It is a tidy line, and it has now been said to every interview panel in the country. What separates a strong candidate is knowing what is genuinely running in NHS services as of 2026, how it was allowed to get there, and which problems the regulator has not solved. Our guide to AI in medicine sets out the ethical frame; this piece is about the machinery underneath it.

Start with a filter that outlives any news cycle. Ask of any AI claim three questions: what task does it do, who checks the output, and what happens when it is wrong? Almost everything genuinely in NHS use answers them the same way. It performs one narrow perceptual or scheduling task, a human remains the decision-maker, and a false result costs time rather than a life, because a person stands between the output and the patient.

Stroke is the clearest case. In a suspected stroke the questions are whether a large vessel is blocked and how much brain is still salvageable, and the answers decide whether someone is thrombolysed, sent for thrombectomy, or neither. Software that reads the CT angiogram, flags a probable large-vessel occlusion and pushes the study straight to the on-call team is not making that decision. It is compressing the minutes before a human makes it. In a pathway where the treatment window is measured in hours and brain tissue is lost by the minute, buying time is the entire benefit — and you can describe that benefit precisely without quoting a single accuracy figure.

Now read the pattern rather than the rows. Every deployed use above is assistive, bounded and audited. The uses that fill headlines — a model that diagnoses unsupervised, a chatbot that manages a patient end to end, a system that decides who is treated first — are not in NHS service, and the obstacle is not computing power or clever enough engineering. It is that nobody has answered who is responsible when such a system is wrong.

Ambient documentation is the interesting middle case. It spread quickly because it touches no diagnosis, only the note — and yet the note is a legal record, a medico-legal document and the thing the next clinician relies on, so national guidance on safe use arrived after the products rather than before them. As of 2026 the honest summary is that the NHS adopts AI fastest exactly where the accountability question is easiest to answer. That is not an accident, and noticing it is worth more at interview than any product name.

Here is the part most candidates have never been told, and it is what makes an answer sound like it came from someone who read past the headline. In UK law, software is not a special category. If a product has a medical purpose — diagnosis, prevention, monitoring, prediction, treatment — it is a medical device, and it is regulated by the Medicines and Healthcare products Regulatory Agency under the Medical Devices Regulations 2002. A cycle-tracking app that only records what you type into it is not a device. The same app, the moment it tells you that you probably have a condition, is one.

Risk classification then sets the scrutiny: the greater the harm a wrong output could cause, the more evidence the manufacturer must produce and the more an independent assessment body is involved before the product may be placed on the market. Notice what is being assessed. Not “the AI” in the abstract, but a specific version of a specific product, with a stated intended use, validated on stated data, for a stated population. Point the same model at a different population, or use it for a purpose the manufacturer never claimed, and you are outside the approval entirely. That is why “is this tool validated?” is a weaker question than “validated for whom, and for what?”

The two-gate logic that governs medicines applies here too, and saying so links this topic to every other NHS question you will be asked. A device may be lawfully marketed and still be something the NHS should not buy. NICE assesses digital health technologies against an evidence standards framework, revised more than once since it first appeared, which scales the evidence demanded to what the product claims: a tool that only supports administration is asked for less than one that shapes a diagnosis. NICE has also used early value assessment to say, in effect, that a technology looks promising enough to use while more evidence is collected — a conditional yes rather than a verdict. If the division of labour between regulators, funders and providers is hazy, our guide to how the NHS works lays out who decides what.

A model learns the distribution it was shown. If a dermatology dataset contains few images of darker skin, the resulting classifier performs worse on darker skin — and it will not announce that. It will simply be confidently wrong more often for some patients than for others. The failure is not malice, or even sloppy code. It is inherited from who was in the data.

The sharper version is the proxy variable. A widely cited 2019 study of a commercial risk-prediction algorithm used in the United States found that it ranked patients by predicted healthcare costs as a stand-in for how ill they were. Because less money had historically been spent on Black patients at the same level of illness, the algorithm systematically under-identified them as needing extra care. Every step in that chain was statistically defensible; the outcome was inequitable. That is what makes bias an ethics problem rather than an engineering one, and it lands squarely on justice in the four pillars of medical ethics. An NHS that deployed such a tool without asking whose data trained it would be automating an existing inequality at scale and calling the result objective.

Which brings you to explainability. Many high-performing models produce an output with no reason a human can follow — the black-box problem. The clinician remains accountable either way, so a tool that cannot show its working leaves them in an unfair position: agree without understanding, or disagree without evidence. Some products mitigate this by showing what drove the flag, highlighting the region of a scan the model weighted most. That is not an explanation, but it is at least an audit trail — a distinction worth drawing out loud.

Two quieter harms sit beside it. Automation bias is the documented human tendency to defer to a machine: to stop looking once the software has said normal, and to doubt a correct reading of your own when the software disagrees. Deskilling is the same effect stretched over years. If trainees only ever see studies a model has pre-sorted, the rarest and most consequential patterns are exactly the ones they may never learn to catch unaided. Both are arguments about how a tool is used rather than arguments against using it, and saying that shows you can criticise a technology without rejecting it.

So: will AI replace doctors? Answer by task, not by profession. AI substitutes well where the work is pattern recognition on standardised data with a knowable right answer — a retinal photograph, a chest film, a rhythm strip. It substitutes poorly or not at all where the work is taking a history from someone who is frightened and has not yet mentioned the thing that matters, examining a body, weighing an atypical presentation that fits no pattern, sitting with uncertainty, breaking bad news, or carrying responsibility for what follows. Accountability is the hard limit: a model cannot be held to account, so a person must be.

The realistic expectation is redistribution rather than replacement. Specialties heavy in image interpretation will change most, and change in medicine has never meant disappearance; the plausible future is fewer hours on first-pass reading and more on the parts no system can absorb, inside a workforce already under the strain described in our piece on NHS pressures and the workforce. Add the honest caveat when you say it: that is a prediction, and predictions about automation have a poor track record. Offer it as a judgement, not a fact.

This topic rarely arrives as a question about NHS policy. It arrives as:

• “Will AI replace doctors?”
• “What are the risks of using artificial intelligence in healthcare?”
• “Should an algorithm ever decide who gets a treatment?”
• “You are a foundation doctor. The software has flagged a finding you cannot see on the scan. What do you do?”
• Follow-ups on responsibility: “Who is to blame when it is wrong?” and “Would you tell the patient that software was involved?”

The depth expected is modest but specific: one concrete deployed example, one sentence on how such software is regulated, and one problem you can argue in both directions. What earns marks is a candidate who can say what a stroke tool actually does, then admit plainly that liability is unresolved. For a station scenario the answer has a reliable shape — the tool is a second opinion, the responsible clinician still examines the patient and documents the reasoning, and a disagreement you cannot resolve gets escalated rather than quietly overridden. AI sits with assisted dying in the small set of topics where panels mark balance, not opinion. Rehearse both out loud; our interview preparation pages exist for exactly that.

Hot topics decay; mechanisms compound. Whichever product is in the news the week of your interview, the questions underneath it will be the same four — is this a medical device, who validated it and on whom, what happens when it is wrong, and who answers for that. Learn those and you can meet a tool that does not exist yet without being caught out.

Work through the rest of the system on our interview reading path, check how the schools you have applied to run their stations on our medical school pages, then say the answer out loud to somebody. The candidates who score on AI are the ones who have already argued both sides of it with another person in the room, and discovered which half of their argument was thinner than they thought.