When a fracture is missed on an AI-assisted radiograph review, who is responsible? The answer, under current UK law and NHS governance frameworks, is clear: the clinician. But the question is worth dwelling on — not because the answer will change soon, but because the implications of that answer are not yet embedded in how most orthopaedic surgeons document AI-influenced decisions, how departments govern AI tool deployment, or how trusts manage accountability when things go wrong.
This is not a theoretical governance post. These are the things you need to know now.
Software as a medical device
In the UK, AI tools used in clinical practice are regulated by the Medicines and Healthcare products Regulatory Agency (MHRA) under the medical device framework, specifically the Software as a Medical Device (SaMD) pathway. An AI tool that assists in diagnosis, triage, or treatment planning is classified as a medical device. The manufacturer bears responsibility for CE or UKCA marking, post-market surveillance, and incident reporting.
What this means for clinicians: if a CE-marked AI tool performs as intended and is used correctly, liability sits primarily with the manufacturer if the tool fails. But “used correctly” is doing a lot of work in that sentence. Correct use requires that you understand what the tool is designed to do, what its stated limitations are, and that you apply it within its intended use case. Using a fracture-detection algorithm designed for tibial fractures on pelvic radiographs is not correct use, even if the interface doesn’t stop you.
NICE and the evidence standards framework
NICE published the Evidence Standards Framework for Digital Health Technologies (DHTs) in 2019, updated since, which sets out what evidence is needed to support NHS adoption of AI tools. The framework distinguishes between tools that inform decisions (lower evidence bar) and tools that drive decisions (higher bar, including requirement for randomised evidence in many cases).
The practical implication: a tool appearing on a procurement list does not mean NICE has evaluated it for your use case. A predictive model trained on US registry data does not necessarily apply to an NHS district general hospital. The framework is a guide for commissioners — individual clinicians still need to interrogate whether the tool has been validated for their patient population.
Documentation when AI is involved
The Medical Defence Union and NHS Resolution have both published guidance making clear that AI-influenced clinical decisions require the same standard of documentation as any other decision — and arguably more, given that the reasoning process is less transparent when a machine learning model is involved.
In practice: if you used an AI risk stratification tool to inform a decision about operative timing or patient selection, document that you used it, what it flagged, and why your clinical judgement aligned with or deviated from the tool’s output. If an AI tool flagged a radiograph as low-risk for fracture and you agreed, write that down. If it flagged high risk and you disagreed, document your reasoning.
This is not bureaucracy for its own sake. It is the contemporaneous record that protects you if the case is reviewed months or years later — and when it comes to AI tools, “the algorithm suggested X” is not a clinical defence unless you can also show that you applied independent judgement to that suggestion.
The governance question at department level
Most NHS orthopaedic departments now use at least one AI tool — a PACS system with an AI triage overlay, a templating platform, or a deterioration monitoring tool. In many cases, these have been procured by trust IT or radiology without formal clinical governance sign-off at department level.
This matters for two reasons. First, you may be using a tool without knowing its evidence base, failure modes, or whether it has been externally validated in a population similar to yours. Second, if a patient is harmed and the case goes to litigation or inquest, the department will be asked whether there was a process for evaluating and monitoring AI tools. A credible answer requires more than “it came with the PACS upgrade.”
The minimum governance standard for AI tools in a clinical department: know what tools are in use, understand what they’re designed to do, have a mechanism for reporting unexpected or concerning outputs, and document that process somewhere.
Three things to do, regardless of seniority
First, know the AI tools you’re using — not technically, but clinically. What is each tool designed to detect or predict? What are its stated limitations? Was it validated in a population similar to yours?
Second, document AI involvement in decisions. Keep it brief — one sentence is sufficient — but make it contemporaneous.
Third, if your department doesn’t have a clinical governance process for AI tools, raise it. You don’t need to run it. But knowing it doesn’t exist and doing nothing is itself a governance failure.
The clinician remains responsible. That position is unlikely to change quickly. The question is whether your practice reflects that responsibility — and whether, if challenged, you could demonstrate that it did.