What Can AI Do Today?
Introduction: AI Is Already in the Building
AI is already becoming a part of healthcare.
AI tools are reading scans in radiology departments overnight. Alerting ward teams to patients who may be deteriorating. Drafting the letters a GP sends to consultants, and the discharge summaries that follow patients home after a hospital stay. In some surgeries, AI systems are helping decide how urgently patients need to be seen.
This isn't a future scenario anymore. AI adoption is happening across healthcare systems worldwide and the pace is accelerating.
Most people, whether they work in healthcare or receive it, haven't had a chance to properly understand what these systems can actually do. The headlines swing between breathless optimism ("How AI Can Beat Cancer") and alarming scepticism ("How worried should you be about an AI apocalypse"). Neither is especially helpful.
This unit cuts through both. We'll look at exactly what AI can and can't do across four areas of healthcare right now: medical imaging and diagnosis, clinical decision-making, administration, and patient-facing tools. As well as a little insight into the speed of acceleration.
That's a more honest and more useful place to start.
AI in Imaging, Diagnosis and Clinical Decision-Making
These are the settings where AI has made its most significant inroads, and where the evidence is strongest.
Reading scans and images
AI is already being used to analyse chest X-rays, mammograms, MRIs, and retinal photographs with accuracy that rivals specialist clinicians in specific tasks.
A 2025 study across 12 sites in Germany involving 463,094 women and 119 radiologists found that AI-supported mammography increased breast cancer detection by 17.6% while recall rates remained noninferior to standard double-reading. Crucially, the AI acted as a safety net: it triggered an alert when a radiologist read a mammogram as normal but the AI flagged it as suspicious. This additional safety-net resulted in 204 breast cancer diagnoses that may have otherwise been missed.
Source: Eisemann, Katalinic et al., Nature Medicine, January 2025
These systems work best as assistants: triaging the most urgent cases to the top of the queue, catching findings that might be missed under pressure, and processing high volumes at speed. What they don't do is replace the clinician's contextual judgement, the ability to integrate a scan with a patient's full clinical picture and history. And if a systems training data is skewed (predominantly from one ethnic group or even scanner type) performance can be meaningfully worse for patients who don't match that profile. This is a documented problem, not a theoretical one.
Spotting deterioration and supporting clinical decisions
Beyond imaging, AI is being looked at in assisting clinical decision-making itself. Early warning systems using consistent monitoring can detect the subtle patterns that signal a patient is deteriorating. Often this is hours before a crisis is visible, giving clinical teams more time to intervene. As shown in a 2025 study, an AI algorithm was able to predict patient deterioration up to 17 hours in advance of adverse clinical outcomes. Scheid, Zanos et al., Nature Communications, July 2025
In primary care, AI is showing similar promise. A real-world study across fifteen clinic sites in Kenya found that an AI clinical copilot reduced diagnostic errors by 16% and treatment errors by 13% - with clinicians appearing to learn from the tool over time. However, despite this, the same evaluation showed no reduction in treatment failures. Of note the tool's developer also co-authored the study. Korom, Robert, et al. arXiv preprint arXiv:2507.16947 (2025).
The risks here are real and worth naming. If a system generates too many alerts, clinicians begin ignoring them, including the critical ones (this phenomenon is known as alert fatigue). And the risk of over-reliance is genuine: clinicians, particularly trainees, may defer to an algorithm rather than developing the underlying judgement that would let them recognise when the algorithm is wrong. This will be touched on in more detail in Unit 3.
If an AI system flagged a patient as high risk and a senior clinician disagreed, who should make the final call? What would the answer be in your work setting or, if you're a patient, what would you want the answer to be?
AI in Administration and Patient-Facing Care
Healthcare runs on documentation. A single hospital admission generates a clerking note, a drug chart, a discharge summary, a referral letter, and a coded record entry. Most of this falls to clinicians, carved out from the time they could otherwise spend with patients. Studies consistently show that doctors (particularly those early in their career) spend a large portion of their working time on non-patient facing tasks.
What AI is doing with clinical paperwork
AI tools can now draft discharge summaries, referral letters, and outpatient correspondence directly from recorded consultations, for a clinician to review and approve rather than write from scratch. Early evidence suggests this can cut drafting time & reduce cognitive burden. Clinical coding involves translating complex notes into the standardised diagnostic codes used for billing and commissioning. It is another high-volume task where AI assistance could reduce both time and error rates.
The critical safeguard in all of this is the approval step. In well-designed systems, the AI drafts and humans sign off. The clinician remains responsible for the documentation. The risk is not that AI will act autonomously, it is that time pressure leads to approvals that are too quick and too shallow. A discharge summary containing a wrong drug name or a missed allergy can harm the patient, and mislead the next clinician who relies on it. AI that makes documents faster to produce must not make errors faster to overlook.
Inform patients and get consent before using AI in their care or recording them. Document their response. Patient agreement does not make public AI tools acceptable; follow local privacy policies.
AI in the Hands of Patients
So far, this unit has focused on AI tools used by healthcare professionals. But AI is also being used by patients, to check symptoms, manage chronic conditions, and navigate a health system that can be difficult to access.
This is perhaps the most personal dimension of AI in healthcare. And it is the one where the gap between potential and risk is felt most directly by individuals.
Tools like Ada Health's symptom checker use AI to guide people through their symptoms and help them decide what to do next. Apple Watch's ECG feature, which can detect atrial fibrillation (an irregular heartbeat that significantly increases stroke risk), has received regulatory clearance in both the US and UK and has prompted real clinical referrals.
These tools extend the reach of healthcare beyond clinic hours and clinical settings: a genuine benefit, particularly for people with complex conditions or access limitations.
The important caveat for patients
Not every AI health tool is regulated the same way drugs or medical devices are. Tools that make a medical claim, such as diagnosing or monitoring a condition, should be registered as medical devices with the MHRA. Many wellbeing and information apps are not, and it is often hard to tell which is which. When you use an AI tool to understand your health - whether it's a symptom checker, a wellbeing app, or a diagnostic aid - it is worth asking: Who built this? What was it tested on? Has it been evaluated by someone other than the company who built it? The below study suggests telling the difference between genuine advice and misinformation is challenging:
A 2025 study presented 300 participants with medical questions answered by either doctors or AI, with the source hidden. Participants were unable to distinguish AI responses from clinician responses - and rated even low-accuracy AI responses as equally trustworthy as doctor answers. They showed equal willingness to follow the advice, including in cases where the AI response was known to researchers to be inaccurate.
Would you want to know if a letter from your GP or hospital had been drafted by AI? And would you trust a symptom checker as much as you'd trust a clinician, and if not, what would need to be different?
Unit 1 Summary
Here is what we covered:
Did anything in this unit surprise you, or shift how you think about AI in healthcare, even slightly? Note it down and take it with you as we move on to understanding how these systems work.
If you have time, the bonus page below gives a sense of just how fast this field is moving.
How Fast Is This Actually Moving?
One of the most disorienting things about AI right now is that it is genuinely difficult to calibrate the pace of change. The field moves fast enough that experts regularly revise their estimates.
This page tries to give you an honest, grounded sense of that pace. Figures as of August 2026.
The benchmark problem
One way researchers measure AI progress is through benchmarks: standardised tests designed to assess capability in a specific area. There are many different types of medical benchmarks, including the US Medical Licensing Examination (USMLE), clinical reasoning datasets, radiology interpretation tasks, and pathology identification challenges.
The trajectory of AI performance on some of these benchmarks is striking.
In 2022, large language models were scoring around 50–60% on the USMLE, roughly the pass mark. By early 2023, GPT-4 was scoring over 85%. By 2024, specialised medical models were performing at or above the level of specialist clinicians on several disease-specific question sets.
Performance on a single benchmark is not the same as clinical performance, a distinction that matters enormously. A model can ace a multiple-choice medical exam while being dangerously wrong in the messy, ambiguous reality of clinical practice. But the trajectory is real, and ignoring it would be a mistake.
What the people building these systems are saying
The leaders of the organisations developing the most powerful AI systems OpenAI, Google DeepMind, Anthropic, Meta have become notably more candid in recent years about the pace and potential scale of what they believe they are building.
Dario Amodei, CEO of Anthropic, wrote in late 2024 that he believes AI could, within the next few years, "compress decades of scientific progress into just a few years" in biology and medicine, potentially running clinical trials autonomously, developing and testing new treatments at a pace no human research system could match.
These are the views of people with obvious commercial interests in projecting confidence. They are not predictions that should be taken as certainties. But the fact that several leaders in competing companies are converging on similar assessments is worth noting.
The healthcare-specific lag
Healthcare tends to adopt new technology more slowly than other sectors, for reasons that are largely good ones. Regulatory approval takes time. Clinical validation across diverse populations takes time. Building the trust of clinicians and patients takes time.
This lag is protective in some ways: it reduces the chance of deploying systems that haven't been adequately tested. It also means that healthcare will not absorb the full impact of AI advances overnight, even as those advances accelerate.
The tension is real. What was approved 3 years ago may need reassessment today. The institutions and processes that govern medicine were not designed for a technology that improves this quickly.
What "fast" means in practice
To make this concrete: researchers who were cautiously discussing the possibility of AI-assisted diagnosis five years ago are now testing AI systems that take autonomous clinical actions. Regulatory agencies that developed frameworks for software as a medical device are now reviewing applications for continuously adaptive AI systems, a capability existing frameworks weren't designed to handle.
This shift is occurring within current institutions and systems, challenging leaders to scale existing operations to match the speed of progression.
How does the pace of change described here make you feel? Is that feeling different depending on whether you're thinking about it as a patient or a professional? What would help you feel better prepared?