"What's Next? The Research Frontier — Geoffrey Hinton on radiologists"

prediction is right, and for once the proposal line filed it correctly and the id is accurate on all three of its parts: it is Hinton, it is about radiologists, and it is 2016. That is rare enough in this canon to be worth saying — rosenblatt-nyt-1958 had to spend its first four paragraphs disowning two-thirds of its own filename, and alexnet-2012 and mycin-1976 both open by reclassifying themselves. This file has no filing to correct.

It has two other things to correct, and they matter more than a kind would.

First, the setting is misremembered everywhere. Every retelling I found calls this "a 2016 talk", "a lecture", "a seminar", or "a conference speech" — as though Hinton had taken a stage, prepared remarks in hand, and delivered a thesis about radiology. He did not. The Creative Destruction Lab's own printed program for the day puts him at 2:45 p.m. on 27 October 2016, in a CIFAR-presented panel titled "What's Next? The Research Frontier", sitting with Yoshua Bengio, Ruslan Salakhutdinov and Richard Sutton. It was an unscripted answer on a four-person panel, opening with "let me start by saying a few things that seem obvious" — the throat-clearing of a man who thinks he is about to say something uncontroversial. That is not a defence of the claim. It is the opposite: the most consequential capability forecast of the decade was an aside, and the reason it travelled is that it was quotable, not that it was argued.

**Second, and this is the fact the secondary literature has entirely lost: the moderator of that panel was Steve Lohr of The New York Times. He is named as such on page 8 of the program. On 14 May 2025** — eight and a half years later — Steve Lohr published the New York Times feature in which Hinton walked the claim back. The reporter who was in the chair when the prediction was made is the reporter who collected the self-grade. I have not seen this noted anywhere, and it is checkable in one page of a public PDF. A canon that exists to keep claims attached to their circumstances should keep that one.

The descent, and the ancestor who is also the author

descends_from: [alexnet-2012], and this is the canon's first case where a prediction descends from its own predictor's result. Hinton is the third author of ImageNet Classification with Deep Convolutional Neural Networks and the third name on the SuperVision entry that won ILSVRC in 2012. The radiology claim is not adjacent to that result; it is a straight-line extrapolation from it. A network had learned to classify a million natural images better than the hand-engineered pipelines that preceded it, in four years the error rate had collapsed, and a chest film is an image. The inference is one step long and the step is the whole error.

That makes this entry the exact companion piece to alexnet-2012, which spends its length demonstrating that almost nothing in the 2012 paper was new — the nonlinearity, the regularizer, the columnar architecture and the convolutional network were all borrowed and credited on the page — and that what was actually new was scale meeting hardware. The 2016 claim inherits that paper's authority without inheriting its modesty. alexnet-2012 is a moment precisely because it demonstrated something on a stated benchmark. This entry is a prediction because it asserted something about a labour market, and the distance between those two verbs is where nine years of error live.

Three ancestors that pull hard and are not listed, with the reason for each.

rosenblatt-nyt-1958 is the closest thing in the canon to a template for this document: a leading neural-network researcher's confident public forecast about what a learning machine would shortly do, amplified by press that did not check it, remembered for a generation as the field's standing embarrassment. Hinton is Rosenblatt's direct intellectual descendant — alexnet-2012 lists rosenblatt-perceptron-1958 as its ancestor, so the line runs through this file. But descends_from in this canon means came out of, and there is no evidence Hinton's claim came out of the 1958 wire story. If anything the inheritance runs the other way: Hinton has spent his career living down the perceptron overclaim, and then made one. That is a finding for the prose, not for the header.

mycin-1976 is the ancestor this entry most deserves and cannot have. That file's subject is "the negative result the genre cannot see" — an expert system that matched or beat infectious-disease faculty in a blinded evaluation and was never used on a patient, with its builders' own written account of why. It is the single most relevant thing in canon/ to what happened after 2016: the performance arrived and the replacement did not, for institutional reasons that were documented in 1984 and re-documented in 2025. But Hinton's claim did not come out of MYCIN. It came out of ImageNet, in apparent ignorance of MYCIN. The absence of that descent is itself the point, and it belongs where the reader can see the argument.

frey-osborne-2013 is the same genre — a displacement forecast, three years earlier, with radiology-adjacent occupations in its ranking — and there is no evidence Hinton read it. Note also that entry's discipline problem, which this one does not share: Frey and Osborne repeatedly and explicitly denied forecasting, and their file has to argue them into the prediction kind over their objection. Hinton made no such disclaimer. He said "within five years" and he said "now", and both are dates.

musk-robotaxi-2019 is not an ancestor because it is three years later, but it is the file this one should be read beside; see below.

What it is

On 27 October 2016, at the second edition of the Creative Destruction Lab's "Machine Learning and the Market for Intelligence" conference at the Rotman School of Management in Toronto, Geoffrey Hinton said this:

> Let me start by saying a few things that seem obvious. I think if you work as > a radiologist, you're like the coyote that's already over the edge of the > cliff but hasn't yet looked down, so doesn't know there's no ground underneath > him. People should stop training radiologists now. It's just completely > obvious that within five years, deep learning is going to do better than > radiologists. [...] It might be ten years, but we've got plenty of > radiologists already.

The remark is three separate speech acts stacked in forty seconds, and they have been treated as one ever since.

A metaphor — the coyote, from the Road Runner cartoons, who remains aloft until he notices the absence of ground. It is a claim about awareness, not about capability: the profession is already doomed and merely hasn't registered it. This is the sentence that gave the quote its half-life, and it is the one part of the statement that cannot be graded, because it asserts an already-true fact about a future that had not happened.

A capability forecast — deep learning will do better than radiologists within five years, possibly ten. Narrow, dated, falsifiable. This is the part Hinton would later say was all he meant.

A policy instruction — people should stop training radiologists now. This is not a forecast at all. It is a recommendation to take an irreversible action in the present on the strength of the forecast, addressed to medical schools, residency programs and students, and it carries a cost if the forecast is wrong. It is also, by a wide margin, the most-quoted of the three.

Joshua Gans — the Creative Destruction Lab's chief economist, on the 1:30 panel that same day, and therefore a witness — records that the line drew no pushback from the room. Bengio, Salakhutdinov and Sutton were on the panel; Lohr was moderating; the audience was several hundred people who had come to Toronto specifically to think about what machine learning would do to work. Nobody argued.

And here is the thing that makes this entry worth the file it occupies. The conference had a printed program. The program opens with a welcome letter from Ajay Agrawal, the CDL's founder, dated 27 October 2016 — the same day, the same building, distributed to the same people. It argues, at length and with a worked example, the exact opposite of what Hinton said hours later. Its example is medical imaging:

> All human activities can be described by five high-level components: data, > prediction, judgment, action, and outcomes. For example, a visit to the doctor > in response to pain leads to: 1) x-rays, blood tests, monitoring (data), 2) > diagnosis of the problem [...] (prediction), 3) weighing options [...] > (judgment), 4) administering treatment A (action), and 5) full recovery with > minor side effects (outcome).

And then, on the next page:

> As AIs improve, the value of human prediction skills will decrease because > machine prediction will provide a cheaper and better substitute for human > prediction [...] However, this does not spell doom for human jobs, as many > experts suggest. That is because the value of human judgment skills will > increase (complements). We'll want more human judgment. For example, when > prediction is cheap, then diagnosis will be more frequent and convenient, and > thus we'll detect many more early-stage, treatable conditions. This will drive > up demand for treatment and emotional support, which is provided by humans.

That is, in October 2016, in print, handed to every attendee: cheap machine prediction raises the value of the complementary human work, expands the volume of diagnosis, and increases rather than decreases demand for the humans. It is what happened. It was not a lucky guess — Agrawal, Gans and Avi Goldfarb had already published the argument that year and would publish it as Prediction Machines in 2018 — and it was sitting in the tote bag of everyone who heard the coyote line.

The correct forecast and the famous one were made in the same room on the same day. Only one of them travelled. Any account of this prediction that presents it as "what everyone believed in 2016" is wrong on the documentary record, and the document is the conference's own program.

Why a reading would cite it

Four concrete occasions, and the first one is live this month.

One: the outer bound of the hedge comes due now. The primary claim was five years from 27 October 2016 — due October 2021. The hedge was "it might be ten years" — due October 2026, roughly ten weeks from this entry's writing. A reading in the second half of 2026 is standing on the due date of the most famous AI displacement forecast ever made, and can grade it against the calendar rather than against a vibe. That is a rare and perishable citation occasion. It will not be available next year.

Two: it is the standing counter-case whenever the "Work and the economy" lens sees a displacement claim. LENSES.md instructs that lens to let "hiring data and shipped deployments outrank surveys and predictions." This entry is the worked example of why the instruction exists. Radiology is the domain where the capability arrived most completely — roughly three-quarters of every AI-enabled medical device the FDA has ever authorised is a radiology device — and where employment, compensation and training positions all went up. When a reading encounters "AI is replacing job X", the useful move is not scepticism in general but this specific comparison: is the evidence better or worse than the evidence for radiology in 2016, which was as good as capability evidence gets and still produced the wrong labour conclusion?

Three: it is the canon's cleanest instance of task-is-not-job. LENSES.md also draws the line between capabilities "claimed vs. demonstrated". This case adds a third column that the lens will keep needing: demonstrated, and deployed, and still not substitutive. The reasons are enumerable and none of them are about model quality — models are built to answer one narrow question per image type, so covering a working day means orchestrating dozens of them; regulatory and liability structures assign the read to a named physician; imaging demand is elastic and rose to absorb the productivity; and image interpretation is only about a third of what the job consists of, the rest being oversight, communication, protocol design and teaching. A reading citing this entry gets that list, which generalises to other professions far better than the headline does.

Four: it belongs beside musk-robotaxi-2019 and is the better half of the pair. The Musk entry grades a vendor forecasting its own product, restated annually, with money riding on it — the easy case, where the incentive to overclaim is visible. This one grades a disinterested scientist with a Turing Award and later a Nobel, forecasting somebody else's industry, with nothing to sell. He was wrong by more. That is the harder and more useful datum, and a base rate assembled only from salesmen would tell a reading nothing it did not already assume.

Disclosure, in the spirit of README rule 7. The rule exists because "no vendor is ever flattered by a project that grades that vendor's predictions." There is no vendor here, but there is a lineage: the backpropagation and deep-learning work Hinton did is upstream of every model this project uses, including whichever model wrote this file. Grading Hinton's forecast is not grading his science, and the entry should not be cited as though it were. The 2012 result was real and is filed as a moment for that reason. What failed in 2016 was an extrapolation from it, made by the same person, in a room where a better one was already in print.

What it got right, and what it got wrong

The claim, in its own words. "People should stop training radiologists now. It's just completely obvious that within five years, deep learning is going to do better than radiologists. [...] It might be ten years, but we've got plenty of radiologists already."

Date made. 27 October 2016.

Date due. October 2021 for the five-year claim; October 2026 for the ten-year hedge. The policy instruction ("now") was due immediately and was actionable from the day it was said.

What actually happened. Take the two graded claims separately, because they came apart, and the entry is worthless if they are not separated.

The policy instruction: wrong, and wrong in a way that cost something

This is the part that failed hardest, and it is the part Hinton has never directly regraded.

By the five-year due date the direction of travel was already clear and by 2026 it is not arguable. In the United States, average radiologist compensation was about $520,000 in 2025 — roughly 48% above its 2015 level — and Medscape's 2026 report puts radiology third among all specialties at about $571,000, having passed cardiology. Training positions did not contract; they hit records. The 2025 match offered 1,208 radiology residency positions, a 4% rise on 2024. The 2026 match offered 1,478 across diagnostic and interventional radiology, up nearly 5%, with 1,241 in diagnostic radiology alone, and filled 97.6% of them. Vacancy is at all-time highs; the Neiman Health Policy Institute projects the shortage persisting for decades on current residency growth, because supply is expanding at roughly 1% a year against imaging demand that is rising faster. At the Mayo Clinic — which has licensed or built more than 250 AI models, concentrated in radiology and cardiology, and staffs a 40-person AI team inside the radiology department — the number of radiologists has grown 55% since 2016. The institution that adopted the technology hardest hired the most humans.

So the profession Hinton told the world to stop training is, a decade later, the third-best-paid specialty in American medicine and cannot recruit fast enough.

And there is evidence the instruction was partly obeyed. A national survey of students at all seventeen Canadian medical schools, run in March 2018 and published in Academic Radiology that November, found 67.7% agreed AI would reduce demand for radiologists, that 48.6% of students who ranked radiology first still reported AI-related anxiety about the choice, and that about one in six students who would otherwise have ranked radiology first would not consider it because of that anxiety. The 2026 match shows the diagnostic radiology applicant pool down roughly 14% over three years even as positions hit records.

The careful statement is this, and I will not put it more strongly: survey evidence establishes that AI-displacement anxiety deterred a measurable fraction of students from radiology in the years immediately after 2016; it does not establish that Hinton's remark caused the present shortage. Radiology workforce shortfalls have several well-documented causes — residency caps, imaging volume growth, an ageing workforce — and disentangling them is not something this entry can do. What can be said is that the instruction was addressed to exactly the population that the survey found deterred, that it was the single most-quoted sentence in that discourse, and that a radiology resident writing in The New Republic in 2025 named it as having "created unnecessary panic that deterred some people from entering radiology."

That is the rarest property a bad prediction can have and the reason this file is worth more than an anecdote: it was a forecast that could act on the world it forecast, in the direction of making itself false, at somebody's expense. musk-robotaxi-2019 has no equivalent — a robotaxi that fails to arrive does not reduce the supply of robotaxi engineers. A reading that cites this entry for "experts get AI timelines wrong" has taken the least interesting thing in it.

The capability forecast: missed its date, then partly landed in a shape the forecast did not describe

At the five-year mark the narrow claim was not close, and there is an unusually clean instrument for saying so. In September 2021, within weeks of the due date, Freeman and colleagues published a systematic review of AI test accuracy in breast-cancer screening in the BMJ. Of 36 AI systems evaluated, 34 were less accurate than a single radiologist, and all were less accurate than the consensus of two or more. The review's conclusion was that the evidence was "a long way from having the quality and quantity required for implementation into clinical practice" and that AI was not sufficiently specific to replace radiologist double reading. Promising small studies, it noted, did not replicate at scale. That is the five-year claim graded at its due date by an independent review in a general medical journal, and it fails.

Afterwards it improved, and honesty requires recording that too.

• April 2022. Oxipit's ChestLink received CE Class IIb certification — the first regulatory authorisation anywhere for an AI application to report medical images autonomously, with no radiologist in the loop. Its scope is narrow by construction: it reports chest X-rays it judges to contain no abnormality, and passes everything else to a human. Sectra announced its acquisition of Oxipit in March 2026 and completed it in April.
• August 2023 onward. The MASAI trial in Sweden — over 100,000 women randomised at four sites between April 2021 and December 2022, AI-supported reading against standard double reading. The interim safety analysis in The Lancet Oncology reported a 44% reduction in screen-reading workload; subsequent publications through 2025–2026 report a cancer detection rate roughly 29% higher, a lower interval-cancer rate, and no increase in false positives. This is a real, randomised, population-scale result, and it is the strongest evidence that the capability half of the claim was pointing somewhere true.

Note the shape of both. MASAI is AI-supported reading — the workload reduction is achieved by replacing one of two human readers, not both, and the gain is measured as a joint human-plus-model system outperforming humans alone. ChestLink's autonomy is real and is confined to the negatives. Neither is "deep learning does better than radiologists." Both are "deep learning plus a radiologist does better than radiologists" — which is what Agrawal's welcome letter said would happen, in the same program, on the same day, before the coyote line was spoken.

Meanwhile the technology diffused enormously without displacing anyone. As of early 2026 the FDA has authorised roughly 1,451 AI-enabled medical devices, about 76% of them in radiology imaging — the most consistent finding in every independent count from 2021 to 2026. As of March 2026, none of them uses a large language model or generative AI. Adoption in practice lags authorisation badly: a 2024 investigation found 48% of radiologists using any AI at all, and a 2025 survey found only 19% of organisations deploying radiology AI reported high success.

Hinton's grade of himself, and an independent one, and they differ

His. Speaking to Steve Lohr for the New York Times feature published 14 May 2025, Hinton said he had spoken too broadly in 2016 and had not made clear that he was talking purely about image analysis; that he was "wrong on the timing but not the direction"; and that most medical image interpretation will end up done by a "combination of A.I. and a radiologist, and it will make radiologists a whole lot more efficient in addition to improving accuracy." Secondary reporting of his recent remarks puts his revised horizon at ten to fifteen years from around 2024. In some accounts he has described the 2016 statement as a "daring and incorrect prediction" — a notably franker self-assessment than "wrong on the timing", but I could not run that phrasing back to a primary source and it is recorded here as unverified.

An independent one, which is harsher on three specific points.

"Wrong on the timing" understates the error by mis-describing what the claim was. Timing errors are venial when a forecast merely announces a future. This forecast issued an instruction to act in the present — "now" — whose entire justification was the five-year horizon. A five-year claim that is a fifteen-year claim does not make the instruction slightly premature; it makes it wrong at the moment it was given, since a radiology residency plus fellowship is six to seven years and the cohort told not to start would have been mid-career before the revised horizon arrived. When the operative verb is "stop", timing is not a detail of the claim. It is the claim.

"Not the direction" is doing more work than it can carry. The direction the field actually took — machine prediction complementing and expanding human judgment, raising volume, raising employment, raising pay — is not a slower version of the direction Hinton named. It is the direction Agrawal put in print at the same conference and explicitly labelled as the alternative to what "many experts suggest". Hinton's 2025 formulation, that AI and radiologists will work together and radiologists will become more efficient, is Agrawal's 2016 position. Adopting your critic's thesis nine years on and filing it as a timing correction is a self-grade this canon should decline to accept at face value, which is exactly why the entry format demands two grades.

The retreat to "image analysis" narrows a claim that was never narrow. "Deep learning is going to do better than radiologists" is arguably about a task. "People should stop training radiologists now" is unambiguously about a profession, and "we've got plenty of radiologists already" is unambiguously about a labour supply. The sentence that did the damage is the one the narrowing does not cover.

The bottom line for base-rate purposes. One of the most qualified people alive, forecasting inside his own specialty, with the best available evidence, about an industry he had no stake in, got the horizon wrong by a factor of two to three and the labour-market direction wrong in sign. He was not corrected by his peers in the room; he was contradicted in the printed program he was handed at the door, by an economist, who was right. That is what this entry contributes to the base rate, and it is worth more than the fact that a forecast missed.

Commonly misused as

Not required for prediction, but this one is misused so heavily in so many directions that the section earns itself. Four misuses, two from each side.

"The radiologist effect" — proof that AI does not take jobs. By far the most common in 2026, and it converts one case into a law. What radiology actually demonstrates is a conjunction: narrow single-finding models, a regulated profession with a named accountable reader, elastic demand for the underlying service, and a job where the automatable task is about a third of the work. Take away any one of those and the case may not transfer. Radiology is evidence about radiology and a hypothesis about everything else. A reading that cites this entry to argue that displacement never happens has done precisely what Hinton did in 2016 — generalised confidently from one domain — with the sign flipped.

"Experts are always wrong about AI timelines." This is the entry being used as a base rate of one. A base rate is what canon/ is accumulating across rosenblatt-nyt-1958, simon-1965, minsky-1970, kurzweil-2005, frey-osborne-2013, musk-robotaxi-2019 and this file, and the interesting question is not whether forecasters miss but which way and by how much. Note that the canon also holds forecasts that were too conservative. Citing one miss as a law is not scepticism, it is the same overconfidence wearing different clothes.

"Hinton was right, just early." This is the self-grade smuggled in as consensus, and it is the misuse this entry most needs to block. It is defensible for the narrow image-analysis claim and indefensible for the sentence everyone actually quotes. Anyone using this line should be made to say which of the three speech acts they mean.

"AI doesn't work in radiology." Flatly false and the easiest to check: about three-quarters of every AI-enabled device the FDA has ever authorised is a radiology device, and MASAI is a randomised population-scale trial showing a material improvement in cancer detection. The technology worked. It did not do the thing that was forecast with it. Those are different sentences, and keeping them apart is most of what this file is for.

Sources

• **Creative Destruction Lab, Machine Learning and the Market for Intelligence, conference program, Rotman School of Management, University of Toronto, 27 October 2016. Primary, read directly (cover and pp. 1–8 of the PDF at creativedestructionlab.com). Source for: the date and venue; the 2:45 p.m. CIFAR panel "What's Next? The Research Frontier" with Bengio, Hinton, Salakhutdinov and Sutton; Steve Lohr of The New York Times as its moderator**; and both quoted passages from Ajay Agrawal's welcome letter, which are transcribed here from the document itself. The letter's closing note credits the argument to Agrawal, Gans and Goldfarb, "Managing the Machines" (2016).
• Hinton's remark itself. Not read in the primary. The Creative Destruction Lab posted the clip to YouTube as "Geoff Hinton: On Radiology" (youtube.com/watch?v=2HMPRXstSvQ); this machine could not retrieve the video page or a transcript — the fetch returned only site chrome. The quotation above is assembled from two independent secondary transcriptions that agree word-for-word on the overlapping span: the European Medical Journal review "Artificial Intelligence in Radiology: An Exciting Future, but Ethically Complex" (emjreviews.com), which carries the opening sentence and the coyote passage, and multiple reports carrying the closing clause "It might be 10 years, but we've got plenty of radiologists already." Recorded as a sourcing weakness: the primary recording exists and was not heard.
• Siddhartha Mukherjee, "A.I. Versus M.D.: What happens when diagnosis is automated?", The New Yorker, 3 April 2017. The article that carried the quote out of the room and into general circulation. Cited from the publication record and from secondary quotation; not read in full.
• Steve Lohr, "Your A.I. Radiologist Will Not Be With You Soon", The New York Times, 14 May 2025. Not read directly — the Times was not retrievable from this machine. Hinton's self-grade quotations ("spoke too broadly", "purely about image analysis", "wrong on the timing but not the direction", the "combination of A.I. and a radiologist" formulation) and the Mayo Clinic figures (radiologist headcount up 55% since 2016; 250+ AI models; a 40-person AI team in radiology) are taken from AuntMinnie's report of it, "Hinton acknowledges mistake in predicting AI replacement of radiologists", and from Radiology Business's coverage. Two outlets, consistent wording.
• Deena Mousa, "AI isn't replacing radiologists", Works in Progress, 25 September 2025. Read directly. Source for: 1,208 radiology residency positions offered in 2025 (+4%); average income $520,000 in 2025, 48% above 2015; 700-plus FDA-cleared radiology AI models at ~75% of all medical AI devices; 48% of radiologists using any AI (2024); 19% of organisations reporting high deployment success (2025); accuracy dropping as much as 20 points outside the training hospital and 38% of 2024 models tested on single-hospital data; 36% of radiologist time spent on direct image interpretation; US imaging utilisation up 60% from 2000 to 2008.
• Joshua Gans, "Radiologists are the canaries in a coal mine", 26 September 2025. Read directly. Gans was the CDL's chief economist and spoke on the 1:30 panel the same day, which makes him a witness: source for the absence of pushback in the room, and for the enumerated reasons the forecast failed (single-finding models, no comprehensive read without a human prompt, regulation, insurer hesitancy, diagnosis as roughly a third of the job).
• Arjun Byju, "The 'Godfather of AI' Predicted I Wouldn't Have a Job. He Was Wrong.", The New Republic, 2025. Read directly. A radiology resident's account: the panic-and-deterrence claim, the ~1,000 FDA-authorised AI devices with at least 76% in radiology, and the report of Hinton's revised ten-to-fifteen year horizon.
• Karoline Freeman, Julia Geppert, Chris Stinton et al. (Sian Taylor-Phillips corresponding; Division of Health Sciences, University of Warwick), "Use of artificial intelligence for image analysis in breast cancer screening programmes: systematic review of test accuracy", BMJ 2021;374:n1872, 1 September 2021, doi:10.1136/bmj.n1872 (PubMed 34470740). Searches ran to 17 May 2021. Abstract and conclusions taken from secondary summaries and the University of Warwick press release; the paper itself was not fetched — PubMed returned only a cookie notice to this machine. Source for the 34-of-36 finding and the "long way from having the quality and quantity required" conclusion. This is the entry's single most important grading instrument and it deserves a direct read on any later pass.
• MASAI trial: interim clinical safety analysis, The Lancet Oncology, August 2023 (PubMed 37541274); screening performance, The Lancet Digital Health, 2024; interval cancer, sensitivity and specificity, The Lancet, 2025–2026. Abstracts and publisher summaries; full texts not read. Source for the 44% screen-reading workload reduction, the ~29% higher cancer detection rate, the lower interval-cancer rate, and the 100,000-plus women randomised at four Swedish sites between April 2021 and December 2022.
• Oxipit, "CE Mark for First Autonomous AI Medical Imaging Application" (oxipit.ai), and AuntMinnie's report of the same. ChestLink, CE Class IIb, launched April 2022, autonomous reporting of normal chest X-rays. Sectra's acquisition announced March 2026 and completed April 2026, per Sectra's own release.
• Bo Gong, J. P. Nugent et al., "Influence of Artificial Intelligence on Canadian Medical Students' Preference for Radiology Specialty: A National Survey Study", Academic Radiology, published online November 2018, in print 26(4), April 2019, pp. 566–577 (PubMed 30424998). Abstract read via publisher summary; full text not fetched. Source for the March 2018 survey of students at all 17 Canadian medical schools, 322 respondents, 29.3% expecting outright replacement but 67.7% expecting reduced demand, 48.6% reporting anxiety among those ranking radiology first, and about one in six deterred outright. The full author list was not verified against the paper; only the first two names are given here, from the publication record.
• Match data: Radiology Business and AuntMinnie coverage of Match Day 2026 — 1,478 total diagnostic and interventional radiology positions (+5%), 1,241 in diagnostic radiology (+3%), 97.6% fill rate (down from 98.4% in 2025), PGY-1 applicant pool down about 1% year on year and about 14% over three years.
• Compensation: Medscape's 2026 Physician Compensation Report, radiology third at about $571,000, and Doximity's 2025 report showing roughly 7.5% growth to $571,749 — both via Radiology Business's reporting; neither report read directly.
• FDA device counts: independent trackers reporting 1,451 cumulative authorised AI-enabled devices through end-2025 (up from ~1,250 a year earlier), 76% in radiology imaging, and no generative-AI or LLM-based device authorised as of March 2026.
• Harvey L. Neiman Health Policy Institute, "Projected US Radiologist Supply, 2025 to 2055", Journal of the American College of Radiology, and the Institute's accompanying press material. Source for ~1% annual growth in radiologist supply and the projected persistence of the shortage. Abstract and press summary only.

Fetch failures, recorded per the rule that a failed search is noted and the work continues. The YouTube recording of the remark (site chrome only); nytimes.com (not retrievable); pubmed.ncbi.nlm.nih.gov (cookie interstitial on every request); radiologybusiness.com (HTTP 403); knowablemagazine.org (HTTP 403); fortune.com's May 2026 retrospective, whose server issued a malformed redirect — its figures (roughly 4,300 active radiologist job listings as of March 2026, averaging 130 days to fill) are reported here at second hand from syndicated copies and are the least well-sourced numbers in this file.