AI in the NHS: rewards, risks, and reality

The NHS must avoid AI’s seductive overdiagnosis trap, writes Professor Christina Pagel, director of University College London’s clinical operational research unit

The government has ambitious plans to use AI across the NHS. Its 10 year health plan for England aims to create a new model of care, with technology at its centre.

To support this, a National Commission has been established to make the NHS “the most AI-enabled care system in the world,” accelerating safe access to AI and shaping a regulatory framework.

Yet success is nowhere near assured. There are risks as well as opportunities and implementation needs to be planned carefully.

Where AI adds real value

Understanding what is meant by AI is vital. A few years ago, the term mainly referred to machine learning (ML) and neural networks, systems trained to recognise patterns, particularly in images.

Today, people using the term AI usually mean large language models (LLMs) such as ChatGPT. The distinction matters. ML excels in pattern recognition, classification and prediction tasks, while LLMs handle language, synthesis, and reasoning.

Detecting more does not always mean treating better. As diagnostic AI expands, the NHS must avoid over-testing and overtreatment

Medical imaging offers the clearest success case so far. ML systems can analyse scans and flag anomalies with remarkable accuracy, making radiology and pathology ideal areas for AI support.

These tools assist rather than replace human experts. That said, ML can also detect harmless irregularities, which can contribute to over-diagnosis.

Given over-diagnosis is already a concern for some conditions such as thyroid or prostate cancers, the use of ML must be considered on a disease-by-disease basis. Detecting more does not always mean treating better, and as diagnostic AI expands, the NHS must avoid over-testing and overtreatment.

The seductive trap will be that the more healthy people who are treated when they don’t need to be, the better outcomes from your programme will look (because you’re diluting the statistics with people who were always going to be fine).

LLMs are emerging as practical tools to support clinicians in real-world decision-making. They can review, summarise and synthesise patient data, identify subtle patterns that might be missed (particularly from text data), and offer alternative explanations, essentially acting as another pair of eyes.

Rather than replacing clinical judgment, they complement it by reducing common sources of error, such as cognitive bias or communication gaps within teams.

By providing input without regard to hierarchy or status, LLMs have the potential to quietly strengthen the way care decisions are made. The key is that LLMs should not be used to replace clinical judgement and expertise, but to enhance it.

AI also promises to relieve administrative burdens. Tools that transcribe consultations, or draft letters, can save clinicians or support staff hours every week. In an overstretched NHS, time is precious, but efficiency must not compromise empathy, the human connection that data alone cannot capture.

Balancing AI and human expertise

Despite progress, hospitals are not fully AI-enabled. While imaging-based AI is in use, most clinicians cannot access advanced LLMs due to legal and ethical restrictions as patient data cannot be uploaded to commercial systems or basic infrastructure issues such as old and slow digital machinery.

Real progress will require secure, ring-fenced models within hospitals, trained on local datasets under strict governance and with upgraded hardware and software. Just as much of an issue is that most UK healthcare settings are understaffed and clinical teams simply struggle to find the time or energy to learn new systems.

Responsible implementation is the key challenge. AI is only as good as the data it relies on; incomplete or biased datasets risk perpetuating inequities.

Historic underrepresentation of women, ethnic minorities, or older adults can skew outcomes, though AI also holds potential to correct these gaps, for example, improving recognition of skin conditions on darker skin or ensuring women’s symptoms are properly assessed.

At one hospital, researchers are currently using camera-based AI to monitor sedated patients (with their prior consent) for signs of pain or delirium.

Early results suggest it can improve comfort and shorten hospital stays, but there may be concerns around surveillance and consent if rolled out more broadly. It is just one example of an exciting but controversial use of AI to reshape the delivery of care and doctor-patient relationships.

Data will never capture the full complexity of healthcare. Lifestyle factors, patient-reported symptoms, and in-person observations are often unrecorded but vital to effective care.

AI can process information faster than any clinician, but human expertise remains essential to interpret results and apply context.

The creation of a National Commission for AI in healthcare is therefore timely. Regulation can build trust, establish ethical guardrails, and accelerate safe adoption. The UK’s ambition to lead the world in AI-enabled healthcare is realistic, but only if innovation and oversight advance together.

AI offers immense promise to improve efficiency, accuracy, and patient care. Yet its success depends on the people who design, deploy, and interpret it.

AI will not replace health professionals, but when implemented responsibly, it can help them work smarter, placing patients at the centre of healthcare.