Article

This is the third in a series of blogs related to IPPR Scotland’s project on ‘Employment, Productivity and Reform in the Scottish Public Sector’ funded by the Robertson Trust.

The Scottish government has huge ambitions for digitising public services. In late November 2025, first minister John Swinney stated that:

“…the potential of technology to improve services and benefit lives stretches across all areas of the public sector. Building this technical foundation will help us develop new ways to grow the economy, end child poverty, improve public services and tackle the climate emergency.”

Indeed, in an earlier speech the first minister seemed to indicate that rapid tech deployment could be an existential issue for Scotland’s public services:

“…we are not going to be able to make the money we have available for public services match the demand for those services unless we ramp up our use of technology.”

These ambitions and concerns are widely shared. Owners of tech firms, media commentators and politicians of all stripes are lining up to argue that technology – especially AI - will radically and positively transform public services by improving quality and lowering costs. We are told that AI will enable fewer people to deliver a level and quality of service that will, at the very least, match current standards.

Let’s take one (now somewhat infamous) example. In 2016, Geoffrey Hinton, commonly identified as one of AI’s founding fathers, asserted that:

“I think if you work as a radiologist, you are like the coyote that’s already over the edge of the cliff but hasn’t yet looked down… People should stop training radiologists now. It’s just completely obvious within five years deep learning is going to do better than radiologists… It might be 10 years, but we’ve got plenty of radiologists already.”

As Deena Mousa notes, Hinton’s claim had an intuitive validity given that ‘digital inputs, pattern recognition tasks, and clear benchmarks predominate’ within radiology.

A decade later AI tools have indeed been widely adopted within the radiology profession. Yet demand for radiologists is up not down. In the US – which we might reasonably expect to be at the bleeding edge of AI deployment - radiology’s vacancy rates were at all-time highs in 2025.

A recent Royal College of Radiologists (RCR) report found that Scotland currently has a 25 per cent shortage of radiologists and a 19 per cent shortage of oncologists. Staff shortages are predicted to worsen through the course of the current decade. Only 6 per cent of (UK) NHS clinical directors have found that adopting AI tools reduces workload. Over half (56 per cent) reported no significant change to workload and, remarkably, over one-third (37 per cent) reported an increase.

If Scotland (or any other country) had followed Hinton’s advice to the letter back in 2016 the outcomes would have been catastrophic. Therefore, it is worth interrogating why he was wrong and whether the experience of radiology offers insights that might help to guide effective public service reform in Scotland.

As a number of recent articles (e.g. Deena Mousa; Sarah O’Connor and John Burn-Murdoch) have argued, it seems that AI developers often have a poor understanding of the services they wish to render redundant. They don’t understand how radiologists actually spend their time. They tend to overestimate the impact on total workload of deploying AI tools on some specific tasks. Therefore, they tend to assume that AI will substitute, rather than complement, existing workers.

And, as we have seen in multiple dimensions in recent years, when technology enhances productivity and lowers costs, this can serve to increase demand for the good or service which in turn increases the demand for labour. This is the Jevons Paradox, an economic concept with which readers are likely to become familiar in coming years.

But if productivity has been rising in radiology it clearly hasn’t been sufficient to reduce the demand for radiologists. Lower unit costs are only one factor contributing to rising demand with others including an ageing population, prevalence of chronic diseases, clinician and patient expectations, advances in imaging technology and more comprehensive screening programmes.

The RCR report cited earlier reflects the current reality for radiologists:

“The data shows that demand for imaging persistently exceeds the rate of workforce growth. Despite their dedication and hard work, radiologists are struggling to keep on top of diagnostic waiting lists because demand for their expertise is growing more rapidly than the workforce itself. Despite radiologists working more productively than ever before, the radiology workforce shortfall has not fallen, because there are not enough radiologists to meet the demand they face”.

It wasn’t just Hinton. No other medical specialty has been more hyped by AI enthusiasts for doctor replacement than radiology or clinical imaging. With over 870 Food and Drug Administration approvals, medical imaging is the single largest target for healthcare AI investments in the US. While a few AI tools are helping some clinicians with specific tasks in niche fields (eg Brainomix helping stroke doctors decide which patients are suitable for treatment or acting as the second reporter in mammography interpretation) AI is far from ready, safe or legal for more widespread roll out.

Contrary to much of the popular commentary, radiology is not “just” pattern recognition. Interpretation of medical imaging – be it X-rays, CT or MRI scans - requires a minimum of seven years post-graduate training after medical school in the UK. Clinical radiologists pull together each patient’s complex medical history, their presenting problems and imaging findings; balancing identifying salient findings with not over-diagnosing incidental observations, unlikely to be of clinical significance.

As already noted, demand for medical imaging has hugely increased over the last decade. With insufficient training positions, trainers and consultants – Scotland has just 388 radiologists when the Scottish government’s modelled need is 863 – patients are waiting longer for their scans and treatment. Increasing numbers of scans are sent to private teleradiology companies, costing £46 million between 2018/19-2022/23. Even as AI companies move from “diagnosing” to “triaging” which scans should be read urgently or helping with language in the radiology report – the rate-limiting step remains availability of the radiologist, a human doctor.

The experience of radiology holds important lessons for public service reform in Scotland. AI has helped to improve treatments and outcomes but, because it doesn’t replace workers, total costs tend to increase, not decrease. Indeed, this is characteristic of the health sector as a whole. New technologies raise costs through the direct cost of purchasing new hardware, software and medicines and also the increase in demand that the new technologies bring. They rarely, if ever, provide perfect substitutes for labour.

The danger is that bold narratives about what technology might deliver in the future can have a real impact on how decisions are made in the here and now. If the Scottish NHS had planned on the basis of Hinton’s claim there would be an even deeper recruitment crisis and many cancer sufferers would be denied the treatment they need and deserve.

Rather than waiting for AI solutions that may never fully deliver, policy should prioritise expanding training capacity and retaining radiologists now to improve access, efficiency, and healthcare system resilience.

(Cindy Chew is Honorary Professor at the University of Glasgow)