Forecasts of the effect of AI on jobs usually start from the tasks a machine can perform. Radiology, the medical specialty most exposed to AI, points to a different margin. This column uses US Medicare data for 2013–2024 to show that as AI tools spread, radiologists did not read more scans and the profession kept growing, but radiology work moved out of small local markets and into multi-state groups. The evidence fits a mechanism in which AI makes expert work done elsewhere easier to check. The costs of AI may fall less on occupations than on places.
How many jobs will artificial intelligence destroy? Most forecasts start from tasks, scoring how much of each occupation’s work an AI system could perform (Eloundou et al. 2024). Recent VoxEU columns show how fragile that starting point is. Exposure scores vary nineteen-fold depending on which AI model does the scoring (Yin 2026); a single index nets out work that AI substitutes and work that it complements (Verschuere and Cameron 2026); and occupations have proved far more durable than task exposure implies (Skoglund et al. 2026, Garicano et al. 2026). Where effects do appear, they show up first in hiring and in small firms (Bonin et al. 2026).
Radiology has become the test case for both sides of this debate. In 2016, Geoffrey Hinton suggested that we should stop training radiologists because deep learning would soon read images better than they do (Langlotz 2019). Ten years later, Nvidia’s chief executive cites the same profession as evidence that AI creates jobs: “the number of radiologists has gone up” (quoted in Verschuere and Cameron 2026). In a new paper (Francesconi and Xia 2026), we argue that AI’s main effect in radiology so far has been on something else: not what radiologists do, but where they work, for which organisation, and ultimately how the boundaries of firms providing radiology services are drawn.
Radiology is the most AI-exposed specialty in medicine. The number of FDA-cleared AI tools that radiologists use while reading images rose from 10 in 2017 to 239 by 2025, and by 2020 about 30% of US radiologists reported using AI (Allen et al. 2021). Every one of these tools assists a radiologist who still signs, and remains responsible for, the final report.
Figure 1 shows what happened to the profession. Between 2013 and 2024 the number of radiologists billing Medicare rose by 16%. Interpretations per radiologist stayed flat through the 2018–19 wave of FDA clearances, fell during the pandemic and have not recovered: by 2024 they were 16% below their 2013 level. The fall reflects fewer fee-for-service patients per radiologist as Medicare Advantage expanded; interpretations per fee-for-service patient rose by about 9%. Neither series shows the jump in throughput that a labour-saving reading technology would produce.
What did change was the organisation of the profession. The share of radiologists in groups operating in five or more states rose from 9% in 2017 to 47% in 2025, and the number of such groups from 79 to 448.
Figure 1 Radiologists and their workload, 2013–2024
To see where the change landed, we compare US local markets that differed, before the recent AI wave, in how much hospital imaging their radiologists interpreted relative to market size, and follow their local radiology billing as FDA-cleared AI spread nationally. The design captures how the link between this exposure and local billing changes as AI becomes available; it does not measure the effect of any single practice adopting AI.
By 2024, a market one standard deviation more exposed had about a fifth fewer radiologists billing locally (Figure 2). The decline begins with the clearance wave and is concentrated in markets that started with fewer than ten radiologists, typically places of 50,000–100,000 people served by five to seven radiologists. Larger markets show no such decline, and some small markets lost local radiologist billing altogether.
Patients in these markets did not stop needing scans: imaging received by local residents shows no relationship with exposure. What moved was the reading. Interpretations of hospital and other facility scans billed from small markets fell to about half their 2013 level by 2024, against a 6% fall in the largest markets. Data on where interpretations are physically performed show the same shift, so this is not merely a change of billing address.
Figure 2 Radiologists billing locally in more exposed markets, relative to 2013
Images could be sent anywhere long before AI arrived. Teleradiology was common by the mid-2000s, but it often supplied preliminary overnight reads, while final reports stayed with local practices. The obstacle was not transmission but trust. A radiology report is a credence good: the hospital buying it cannot easily assess its quality, even after the fact (Arrow 1963, Darby and Karni 1973, Dulleck and Kerschbamer 2006). That makes it costly to buy reports from radiologists outside the organisation.
Our model locates AI’s role here. Tools that triage scans, flag abnormalities, and automate measurements also generate information that can be used to check reads done elsewhere. When checking becomes cheaper, hospitals can buy interpretations from providers that pool demand across many sites. Pooling is most valuable where staffing is indivisible, since a small hospital needs radiologist cover whatever its volume, and the fixed cost of verification favours large groups with many clients. As earlier work on information and monitoring technologies has shown, technology can change how firms are organised and where their boundaries lie without changing the work itself (Baker and Hubbard 2004, Van Reenen et al. 2013).
Three pieces of evidence favour this reading over a simple productivity story.
A calibrated version of the model reproduces the decline across small markets and implies that about a third of interpretation work has been reorganised across markets. Yet this saves only about 3% of aggregate reader capacity. Large organisational change can coexist with small aggregate efficiency gains.
Figure 3 Share of specialists in groups operating in five or more states
Because the profession grew, national numbers hide the local losses. Of radiologists billing in 2017 from a single location in a market with fewer than ten radiologists, 30% had stopped billing Medicare by 2024, against 21% in the largest markets; age explains only part of the gap. Those who stayed billed fewer interpretations. Procedures that need a radiologist on site, such as biopsies and drainages, also declined in exposed small markets, and other clinicians did not fill the gap. Claims data cannot tell whether patients travelled or went without.
These places matter more than their share of radiologists suggests. The 431 US markets that began with fewer than ten radiologists account for about 5% of the profession. Together with markets with no local radiologist and rural areas, they contain two in five US hospitals and three in four critical-access hospitals, and serve about one in six Americans.
First, task-based exposure indices miss this margin. Technologies with similar aggregate efficiency gains can have very different effects on where expert work is done, and those effects appear in thin local markets well before they appear in national occupational totals. Baldwin (2020) expected cheaper remote work to bring ‘telemigrants’ from abroad; in radiology, the same force moves expert work from small markets to large ones within a country.
Second, European health systems face the same pressures. The UK is short of 32% of the consultant radiologists it needs, with the shortfall averaging 43% in small acute hospitals against 30% in teaching hospitals. In 2025 the NHS spent £241 million on teleradiology companies, which by clinical directors’ estimates report around a quarter of its CT and MRI scans (Royal College of Radiologists 2026). If AI lowers the cost of checking work done elsewhere, it will speed up this shift. The question is then who captures the gains from pooling – public imaging networks or private platforms – and whether small hospitals keep both the capacity to check the reads they buy and the specialists that on-site procedures require.
Our evidence has limits. The design identifies changes associated with predetermined exposure, not the causal effect of a practice adopting AI; we cannot fully separate better verification from other forces that ease remote work; and we do not observe radiologists’ earnings or patients’ health. But the labour-market effects of AI clearly cannot be read off the tasks machines can perform. They also depend on how AI changes the organisations through which expert work is produced, checked and delivered.
Source : VOXeu
Firms’ access to external finance differs widely across Europe. This column uses data from the…
Whether inflation can ease public debt burdens is an increasingly important question as debt sustainability…
Senders now have more ways than ever to move money across borders, yet costs remain…
An increasing number of central banks have begun disclosing reserve currency shares. This column introduces…
In public discourse on fiscal tensions across OECD countries, the potential contribution of higher economic…
The market for rare earth elements is relatively small in value, but shortages can lead…