NHS AI diagnostics don’t have an accuracy problem. They have a liability problem.

Just 28% of GPs use AI tools at all, and only a fraction use them for an actual clinical decision rather than writing up notes, according to the Royal College of General Practitioners and the Nuffield Trust’s survey of 2,108 GPs, published in December 2025. The obvious explanation is that the technology isn’t trusted yet. A separate Corti and YouGov survey of UK healthcare professionals found fear of errors is the biggest reservation clinicians name about clinical AI, at 62%, well ahead of the 44% reported in France, and fewer than a quarter feel comfortable using AI tools at all. That’s a real, specific worry, and it deserves to be taken at face value before reaching for a sharper explanation.

Look at the technology itself, though, and accuracy stops looking like the constraint. The US Food and Drug Administration’s list of AI and machine-learning-enabled medical devices has grown from six authorisations in 2015 to 295 in 2025 alone, roughly 49% annual growth through most of that decade and 1,451 devices cleared cumulatively. Closer to home, an NHS AI Diagnostic Fund-backed evaluation of an AI chest X-ray triage tool for suspected lung cancer found 95.87% sensitivity for flagging abnormalities needing urgent review. Clinicians aren’t opposed to the principle either: Health Foundation polling of 1,292 NHS staff in mid-2024 found 76% support using AI in patient care. Whatever is slowing NHS adoption, it isn’t a shortage of capable tools or clinicians who reject AI outright.

What clinicians say they’re actually worried about

Ask non-users why they haven’t adopted AI, and the answer clusters elsewhere entirely. In the Nuffield Trust and RCGP survey, 89% of GPs who don’t use AI named professional liability and medico-legal risk as a key concern, and 80% of GPs who do use AI said the same, meaning experience narrows the worry but doesn’t remove it. A further 88% of non-users cited a lack of regulatory oversight, both figures ahead of fear of clinical error itself. That isn’t a one-off finding: NHS England’s own workforce horizon-scanning research into clinician confidence in AI reaches the same conclusion independently, naming legal accountability as the central barrier to trust, ahead of accuracy or usability. Liability, not competence, is the number clinicians keep returning to.

Who actually carries the risk

That number holds up once you see who owns it. General Medical Council guidance is unambiguous: doctors are responsible for decisions they take using AI, whether or not the tool that informed the decision was flawed, poorly validated, or deployed without adequate training. The Medicines and Healthcare products Regulatory Agency has confirmed it won’t publish a dedicated AI medical device framework until 2026; until then, the only live testing ground for new rules is AI Airlock, a sandbox still working through its first cohort of pilot projects. Medical Protection, the doctors’ defence body, put the position bluntly in a June 2026 report calling for legislative reform: UK product liability law doesn’t clearly classify AI systems as products, so the firms that build and sell them are largely shielded from the rules that would normally apply to a defective device, leaving the clinician, and by extension the NHS, to absorb the claim instead. Medical Protection describes the gap between the law and the technology as feeling “less like a step and more like a widening gulf.”

That gulf sits on top of a liability bill already large and rising before AI arrived. NHS Resolution’s new clinical negligence claims climbed from 11,667 in 2019/20 to a peak of 15,078 in 2021/22, settling at 14,428 in 2024/25, with payouts reaching £3.1 billion that year and future liability provisions standing at £60.3 billion. Set that beside the roughly £144 million the government has committed to AI diagnostics through the AI Diagnostic Fund and its predecessors since 2023: NHS Resolution paid out more than 21 times that figure in clinical negligence in a single year. A clinician weighing an AI second opinion isn’t stepping into a small, contained risk. They’re stepping into a system that already pays out billions annually, with no settled rule for how much of a future claim lands on them personally versus the vendor who built the tool.

Why the caution is rational, not excessive

Medical Protection wants Parliament to legislate AI systems as products subject to strict liability, spreading responsibility across developers, suppliers and deploying trusts rather than concentrating it on the clinician who accepted the AI’s suggestion. A formal consultation is expected in the second half of 2026, feeding the Law Commission’s wider review of product liability for digital products. Whether that produces a working framework quickly is doubtful: core UK medical device regulations were meant to apply from summer 2024 under a 2022 timetable, didn’t take effect until July 2025, and the AI-specific rules on top have already slipped into a 2026 publication date. A consultation opening in late 2026 is unlikely to produce settled legislation on any timescale that changes how clinicians are exposed today.

None of this makes NHS clinicians precious about new technology. It makes them accurate readers of their own incentives. A doctor who defers to a well-validated AI tool, and is later found to have relied on a flawed one, currently owns that outcome alone; contract terms shield the vendor, not clinical negligence law. Until legislation assigns a proportionate share of that risk to whoever built and sold the model, individual caution is the rational response, and adoption will track the pace of reform, not the pace of the technology. Any leadership team building a business case for clinical AI on accuracy data alone is building it on the wrong evidence. The real gating question was never whether the model works. It’s who is contractually and legally exposed when it doesn’t, and that needs settling before the model reaches a ward.

Sources

• Royal College of General Practitioners and Nuffield Trust, “How are GPs using AI? Insights from the front line” (December 2025)
• Corti and YouGov survey of UK, French, German and Danish healthcare professionals, reported by Digital Health (January 2025)
• The Health Foundation, “Majority of NHS staff support using AI in patient care, major polling finds” (July 2024)
• US Food and Drug Administration, Artificial Intelligence and Machine Learning (AI/ML)-Enabled Medical Devices list
• NHS AI Diagnostic Fund-backed evaluation, chest X-ray abnormality triage study, medRxiv (2024/2025)
• General Medical Council, “Artificial intelligence and innovative technologies” guidance
• NHS England / Health Education England, DART-ED horizon scanning: understanding healthcare workers’ confidence in AI
• Medicines and Healthcare products Regulatory Agency, AI Airlock regulatory sandbox
• GOV.UK, Life Sciences Sector Plan (2026 dedicated AI medical device framework commitment)
• Medical Protection Society, “Closing the AI liability gap: AI, safety, and the case for legislative change” (June 2026)
• NHS Resolution, Annual Report and Accounts 2024/25
• Department of Health and Social Care, AI Diagnostic Fund announcement (June 2023)