The mechanism is real and the health care examples are well chosen. Phacoemulsification made cataract surgery fast and safe and volumes rose. Joint replacement followed the same curve. Genetic testing and MRI both got cheaper and both got used more. Khullar is careful to add the condition that actually governs the result: the effect depends on price and on who captures the value, and systems sold at prices that let vendors keep most of the surplus will not stimulate much new demand.

Jevons requires that the buyer face the price. A mill owner bought his own coal. American patients do not buy their own care, and the entity that does is working from a budget rather than a demand curve. Between prior authorization, network design, global budgets, and capitation, third party payers exist in part to prevent exactly the elasticity Jevons describes. Cheaper production of care does not automatically become more purchased care when purchasing is rationed on purpose.

The dialysis world is the clean counterexample. Under a bundled per treatment payment, and more so under a capitated total cost of care arrangement, the payer's stated objective is fewer encounters, fewer admissions, fewer sessions in a chair. Give that system a tool that reduces the labor per patient and it does not buy more labor. It books the savings. Jevons runs backwards under capitation, which is roughly a third of Medicare and a growing share of everything else.

The framing is correct and usefully explains why health care and education got expensive while televisions got cheap. Wages in labor intensive sectors have to keep pace with wages in sectors that actually gain productivity, or the workers leave.

The example he borrows works against him. A string quartet is expensive because four people have to play for forty minutes. Recording did not fix that. It went around it, and the result was that recorded music became ubiquitous and inexpensive while the number of people earning a living as performing musicians did not grow to match. Consumption exploded. Employment did not follow it up. That is the honest version of the analogy and it is a contraction story, not an expansion one.

There is a second problem underneath. Baumol's cost disease operates on wages, not on headcount. If AI cures health care's cost disease, the mechanism by which it does so is slower wage growth for clinicians. You can hold headcount flat, or even grow it, and still have that be the bad outcome for the people in the room.

The observation is accurate. There are more radiologists practicing in the United States than a decade ago, imaging volume has grown faster than the workforce, and the field that was supposed to be automated out of existence is instead short staffed. ACR Harvey L. Neiman Health Policy Institute figures put radiologist growth at 17.3 percent from 2014 to 2023.

The supporting citation is the AJR Expert Panel narrative review by Rozenshtein and colleagues, and that paper is not an optimism paper. According to PubMed, its own abstract describes "a mismatch between the demand for radiologist services and the current size of the radiologist workforce," along with "dissatisfaction, turnover, and burnout among radiologists" and pressure on resident education. Khullar reads "larger workloads" as evidence of healthy Jevons expansion. The panel reads the same fact as strain. Both readings fit. Only one of them is a reason to feel good.

The deeper issue is the counterfactual. Radiologist headcount grew through a period in which clinical AI was mostly pilots and a small number of narrow FDA cleared tools. Reading the last decade as a verdict on autonomous agents is reading a trend line from before the intervention started.

This is the strongest section and it is the one clinicians will recognize. Autor's task framework is the right tool, the surgeon example is well chosen, and the list of skills that resist decomposition (interpreting results, managing uncertainty, consulting colleagues, negotiating treatment plans, motivating behavior change, leading teams) is a fair description of what the job actually is on a Tuesday. The O-ring point is sharp: as systems get more expensive and more effective, small mistakes cost more, and a missed diagnosis matters more when a cure exists.

O-ring theory is symmetric and he only runs it one direction. If the value of careful human oversight rises, that argues for more oversight per unit of output. It argues equally well for fewer, more expert humans supervising much more output, which is the same technology producing wage polarization instead of headcount growth. Kremer's original paper is about matching high skill workers with each other and paying them more, and it comes with a tail of workers who no longer clear the bar. That half is missing here.

The honest hedge is in the essay already, two paragraphs from the end, and it deserved more room: "some types of health care jobs could be replaced."

Clean. No patient appears in the Perspective and none appears in this review. No clinical encounter, no identifiable detail, nothing learned in a position of medical trust. Nothing here requires a permission note.

Ravi Shankar, Clinical Research and Innovation Office, Tan Tock Seng Hospital, National Healthcare Group, Singapore, writing in The American Journal of Medicine (2026). The title is "The Jevons trap: When artificial intelligence in healthcare creates endless work."

Shankar accepts the Jevons mechanism and rejects the conclusion. Induced demand does not arrive as new colleagues. It arrives as more inbox, more documentation to review, more alerts to adjudicate, more imaging to read. Efficiency gets absorbed at the point where it was created. This is the most useful counterpoint because it does not argue with Khullar's economics at all. It argues about who receives the expanded demand, and the answer is the clinician who is already there.

Heathcote Ruthven and Christoph Agten, European Journal of Radiology Artificial Intelligence (2025), "Perspective: AI productivity will not benefit employed radiologists." Open access.

Their frame comes from James Bessen: automation reliably shifts value from labor to capital even when it does not reduce employment. They anchor on the Swedish MASAI screening trial, where AI supported reading cut radiologist workload by 44 percent, and they predict the surplus flows to employers, vendors, and private equity owners rather than to salaried radiologists. Their advice is blunt and worth repeating to any employed physician: get equity in the practice, specialize into something automation resistant, or plan a pivot. Korchi and D'Anna published a reply in the same journal in December 2025, so the disagreement is documented on both sides.

Curtis P. Langlotz, Stanford Institute for Human Centered AI, "The Effect of AI on the Radiologist Workforce: A Task-Based Analysis," medRxiv preprint, December 2025. Not yet peer reviewed.

This is the sharpest one, because it uses the same task decomposition Khullar leans on and comes out the other side. Langlotz projects a 33 percent reduction in radiologist hours worked over five years, with a range of 14 to 49 percent, driven mostly by report generation and study triage. He still concludes that job loss is unlikely for the foreseeable future, because imaging volume keeps growing against a static workforce. Note what that means. The optimistic reading of the optimistic case is that radiology stays fully employed by burning its entire productivity windfall on the backlog. That is Shankar's Jevons trap with a number attached to it. Langlotz is also the source of the line everyone quotes, from RSNA 2017, that radiologists who use AI will replace radiologists who do not.

Jeff Levin Scherz, MD, MBA (Managing Director, WTW; Harvard Chan School), summarizing economist David Cutler on his Employer Coverage newsletter.

Cutler's prediction is that AI enables substitution of less expensive clinicians for more expensive ones rather than a reduction in total clinicians. That is technically consistent with Khullar's thesis and unpleasant in a way Khullar does not address. The workforce can grow while physician roles are the part that gets squeezed. Levin Scherz adds two conditions worth carrying forward: models trained on historical practice reproduce historical inequity, and the Medicare Advantage prior authorization algorithm episode is the available case study in what happens when automated decisions run without meaningful human review.

A fifth, more polemical position exists and is worth reading with your guard up. Kanav Jain's "Algorithmic Austerity" argues that health care AI is primarily an instrument of financial extraction and labor control, and cites an AI driven staffing system that reportedly cut nurse staffing by 30 percent at a large hospital chain. I have not verified that figure, and the piece is an essay rather than research, but the structural claim (that deployment decisions are made by the people who capture the savings) is the same claim Ruthven and Agten make with citations.