Guest blog by Dr Cameron Black
Our recent survey of UK occupational health (OH) professionals found that 60% have experimented with AI tools mainly for documentation, research, and administration. But only 12% use AI in clinical decision-making. The barriers are not technical, they are human: data security, regulatory uncertainty, and professional autonomy. We also found that 80% of OH professionals agree that AI cannot replace expert judgement in complex case management (yet!). That is not Luddism but practical realism.
At an Oxford AI in Healthcare event this year, a doctor described a trial of an AI tool for reading CT head scans. The algorithm performed well, identifying brain injuries faster than radiologists. But when they measured patient turnaround times, nothing changed. The tool worked, but the pathway around it did not. This is the sociotechnical gap, and it is the central challenge facing occupational health (OH) as we integrate artificial intelligence into practice.
Jessica Morley's framework for understanding risks of AI in healthcare is essential reading. She identifies three categories:
- Epistemic risk – AI gives wrong or biased answers. An AI trained on corporate data will fail for gig economy workers. If we do not know what data was used to train a tool, we cannot trust its outputs.
- Normative risk – AI changes practice for the worse. When AI automates routine tasks, there is a real risk of deskilling. If clinicians stop taking detailed histories because "the AI will capture it," we lose the cognitive processing that is essential to good clinical reasoning.
- Traceability risk – No one is accountable when AI fails. If an AI-assisted fitness-for-work decision leads to harm, who is liable? The employer? The OH professional? The vendor? Until this is clear, cautious OH physicians and OH professionals will rightly… hesitate.
Data sovereignty
The NHS has given Palantir access to identifiable patient data for the Federated Data Platform. Whatever your view of that decision, it raises a question every OH professional should ask: where does our data go when we use third-party AI tools? If the answer is "servers outside the UK," that is a UK GDPR problem. Before any AI procurement, ask: where is our data stored? Who owns the anonymised data used to train the model? Can the vendor guarantee that data never leaves UK jurisdiction? If they cannot, do not proceed.
Implementation is the bottleneck
Jessica Morley (2025) offers a simple, way to understand what makes up an AI system.
- Data is the ingredients: the raw input that determines what AI learns from. If your ingredients are poor, biased, or unrepresentative, nothing that follows will be safe or trustworthy.
- Algorithm is the recipe. It is the learning method – the set of instructions that finds patterns in the data. A good recipe can only work with the ingredients it is given. It cannot compensate for poor inputs.
- Model is the meal. It is the finished product that applies what has been learned to complete tasks – diagnosing a condition, predicting return-to-work, drafting a report.
You cannot fix a bad meal by tweaking the recipe alone. You must go back to the ingredients. In AI terms, if your data is flawed, no number of clever algorithms or fine tuned models will produce safe, equitable, or reliable outcomes.
One example data metric is ethnicity. The NHS Ethnicity Recording Improvement Plan reveals persistent and, in some cases, worsening gaps in ethnicity data across care settings. In 2023/24, the proportion of ethnicity data missing from NHS records ranged between 18% and 32% depending on the setting. Inpatient and emergency care providers recorded valid ethnicity codes in approximately 84–86% of cases. Outpatient services fared worse at 77%, and community services at 72%. The independent sector performs significantly worse: valid ethnicity was just 52% for admitted patient care, 52% for emergency care, and only 31% for outpatient services. Since 2020, completeness has declined across most major data sets, including admitted patient care, emergency care, and outpatient services. Consistency across data sets is also poor: among individuals appearing in more than one NHS data set, only 44% had the same ethnicity recorded each time. 26% had a different ethnicity recorded each time they appeared, and 31% had matching records with others showing a different ethnicity.
Incomplete data means the NHS cannot reliably identify who is being left behind. It also means AI and digital tools trained on unrepresentative data risk embedding racial bias, compounding rather than reducing inequalities. As the plan notes: "Incomplete data risks embedding dangerous biases in digital tools (for example, racial bias in a clinical algorithm), which can compound rather than reduce health inequalities." (NHSE, 2025)
Before deploying AI ask: Is the referral pathway ready? Are staff trained? Is there a contingency plan when the AI fails? Have workers and managers been consulted? Is the data robust enough to provide meaningful outputs with employees and employers at the centre?
Without these AI will fail, not because the technology is poor, but because the human and organisational infrastructure is missing.
Occupational health – a call for action
We need to develop ongoing and evidence informed guidance. Do not wait for regulation to be imposed. Build the ethical frameworks, the procurement standards, and the training resources that will make AI safe for the workers we serve. The question is not whether AI will change occupational health, for which it will and has. The question is whether we will shape that change or have it shaped for us.
Dr Cameron Black specialises in integrating technology with occupational health, designing AI-driven ergonomic solutions and telehealth programmes that empower individuals to manage musculoskeletal conditions and reduce workplace injury. Cameron is a founder member of the AI in OH collaboration group.
References
NHS England. (2025) Ethnicity Recording Improvement Plan. Available at: https://www.england.nhs.uk/long-read/ethnicity-recording-improvement-plan/ (Accessed: [22nd May 2026].
Karpathakis, K., Morley, J. and Floridi, L. (2024) 'A justifiable investment in AI for healthcare: aligning ambition with reality', Minds and Machines, 34(4), p. 38.
Morley, J. (2025) When Data Turns Dangerous: The Risks of Bias and Misuse in Healthcare AI. New Haven, CT: Yale Digital Ethics Center.
Morley, J. and Floridi, L. (2020) 'The limits of empowerment: how to reframe the role of mHealth tools in the healthcare ecosystem', Science and Engineering Ethics, 26(3), pp. 1159-1183.
Morley, J. and Floridi, L. (2024) 'The ethics of AI in health care: an updated mapping review', SSRN Electronic Journal. doi: 10.2139/ssrn.4987317.
Morley, J. et al. (2025) 'Global health in the age of AI: charting a course for ethical implementation and societal benefit', Minds and Machines, 35(3), p. 31.