xychart-beta
accTitle: Diagnostic radiology advanced (PGY-2) positions offered and filled in the Main Residency Match, 2022 to 2026
accDescr: Line chart. Diagnostic radiology PGY-2 positions offered and filled in each Match from 2022 to 2026. Offered rose from 997 to 1,083. Filled rose from 996 to 1,066. The two lines are nearly equal every year.
title "Diagnostic radiology PGY-2 positions, NRMP Main Match"
x-axis [2022, 2023, 2024, 2025, 2026]
y-axis "Positions" 900 --> 1120
line "Offered" [997, 1006, 1017, 1057, 1083]
line "Filled" [996, 1006, 1016, 1042, 1066]
9 Will AI take your job? Workforce and economics
AIM-3 · AIM-7 · about 90 minutes
NoteWhat you’ll learn
By the end of this chapter you will be able to:
- Say what was predicted for radiology in 2016, what the original data (match tables and workforce studies) show happened through 2026, and why “he was just early” is not a good enough explanation.
- Ask four questions of the specialty you are entering (does the job break into tasks, does the tool work outside the hospital that built it, who is liable, and does demand grow), and grade a chatbot’s forecast about that specialty with the same four questions.
- Use the story of computer-aided detection in mammography to explain why a tool being adopted and a tool improving outcomes are separate questions.
- Explain why a tool can cut work inside one hospital while raising cost across the whole system, and describe the “bot wars” pattern.
- Read the Bureau of Labor Statistics projections for three health-information jobs and explain why jobs that sit next to each other move in opposite directions.
- Name the task in your intended specialty that a machine could most plausibly take over, estimate what fraction of the job it is, and say what would change your estimate.
Time. About 25 minutes of reading, 20 minutes of watching, 35 minutes of doing, and 10 minutes for the check and the standing question. You need one chatbot and a browser. Your institution’s ChatGPT Edu or Microsoft Copilot account is enough; swap in whatever your own school licenses. Nothing installs, nothing asks for a phone number, and nothing in this chapter touches a patient.
9.1 The year everyone says radiology is over
It is September of your fourth year. You have your radiology letters, your program list, and a spreadsheet of interview dates. At a family dinner, an uncle who reads about technology asks why you would go into the one specialty the “godfather of AI” said should stop training people. He has the quote on his phone. In 2016 Geoffrey Hinton told a Toronto audience that people should stop training radiologists now. Within five years, he said, deep learning would obviously do the job better. Hinton would later share a Nobel Prize for the neural networks that make modern AI work.
Your uncle is not wrong that Hinton said it. He is not wrong that Hinton knows more about deep learning than anyone at the table. The question is what happened next. Does the way the prediction failed tell you anything about the job you are about to spend a decade training for? Not going into radiology? The same uncle has a version of the question for every specialty, and so does every chatbot. This chapter is about how to answer it with something better than a feeling.
9.2 Why this matters
You are making a thirty-year decision with a spreadsheet and a feeling. Nobody can tell you how your specialty will look in 2056. But the most confident, most credentialed prediction that a medical job would disappear has now had its full time to come true. It failed in a specific way that you can learn from. A forecast that has failed is more useful than one that has not been tested yet. You can examine each assumption and see which ones were wrong.
There is a second reason, and your co-residents will not think of it. The people whose jobs are already changing are not physicians. Medical transcription is shrinking. Medical coding is growing. Both jobs are done close to where you will work. A resident who understands why is more useful to the team than one who is only relieved about their own position.
9.3 How it works
9.3.1 What actually happened
The prediction was five years. Ten years on, the National Resident Matching Program’s own tables show what happened in Figure 9.1. Diagnostic radiology offered more advanced (PGY-2) positions in 2026 than in any of the previous four years, and filled 98.4 percent of them. There were 1,741 applicants who ranked the specialty for 1,083 positions.1
The other workforce numbers agree. The Neiman Health Policy Institute counted 37,482 radiologists serving Medicare patients in 2023. It projected a workforce 25.7 percent larger by 2055 even if residency positions stop growing, and 40.3 percent larger if they keep growing (Christensen, Parikh, et al., 2025). Attrition, the rate at which working radiologists leave clinical practice, more than doubled between 2014 and 2022, from 1.1 to 2.5 percent a year. The authors treat that as one cause of a shortage, not a surplus (Christensen et al., 2026). Doximity’s 2026 pay survey put average diagnostic radiology pay at $609,684 for 2025, up 6.6 percent in a year when physician pay overall rose about 2 percent.2
NoteThe subspecialty exception
One number keeps this from being reassurance about every part of radiology. Count the radiologists who bill at least half their work for children. That pediatric radiology workforce shrank from 2,190 in 2016 to 2,032 in 2023, from 6.4 to 4.6 percent of all radiologists, while demand for imaging rose (Morales-Tisnés et al., 2025). That is not automation. It is the job market of one subspecialty moving on its own, apart from its parent specialty. A student choosing a fellowship is making a more specific decision than the overall numbers describe.
In May 2025 Hinton told the New York Times that he had spoken too broadly in 2016. He said he had meant image analysis rather than the whole job. He now expects most image interpretation to be done by “a combination of A.I. and a radiologist.”3 Here is the prediction as it was covered when it was new, and as it reads now.
The New Yorker · magazine
A.I. Versus M.D.: What happens when diagnosis is automated?
Siddhartha Mukherjee · 3 April 2017
The New Republic · essay
The “Godfather of AI” Predicted I Wouldn’t Have a Job. He Was Wrong.
Arjun Byju, radiology resident · 25 October 2024
The New York Times · technology
Your A.I. Radiologist Will Not Be With You Soon
Steve Lohr · 14 May 2025
Radiology Business · trade press
Match Day 2026: Radiology programs offer more positions than ever, but applicant pool declines
March 2026
Read the fourth clipping twice. “More positions than ever” and “applicant pool declines” are in the same headline. The headline does not prove why the pool shrank. But if students are hearing the prediction and acting on it, even as the job market moves the other way, that is a harm the prediction did, and the people who believed it were the ones harmed.
9.3.2 Why “he was just early” is not enough
The easy explanation is that the prediction was only late. Hinton’s own 2025 correction is more interesting. He says he meant a task, image analysis, and said a job. That is not a timing error. The most useful explanation of the failure has four parts, each built into how medical work is organized. Each one is a question you can ask about any specialty. The worksheets later in the chapter call them the four factors.
1. Task decomposition: does the job break into tasks a machine can be tested on? A job can be automated only as far as it breaks into tasks a machine can be measured on. In 2012, observers followed fourteen staff radiologists through their workdays at three hospitals with a stopwatch. They found that 36.4 percent of the radiologists’ time went to interpreting images (Dhanoa et al., 2013). The rest went to protocolling requests (deciding how each scan should be done), supervising studies, procedures, talking with other physicians, and direct patient care (Figure 9.2). A tool that read every image perfectly would leave nearly two-thirds of the day untouched. It would not sign the report either.
pie showData
accTitle: How fourteen hospital radiologists spent their time in 2012, from a time-and-motion study at three Vancouver hospitals
accDescr: Pie chart of a radiologist's day. Image interpretation 36.4 percent, clinical work other than reading images 43.8 percent, other clinical 7.5 percent, nonclinical 12.3 percent.
title "A radiologist's day (percent of time)"
"Image interpretation" : 36.4
"Clinical work other than reading images" : 43.8
"Other clinical" : 7.5
"Nonclinical" : 12.3
2. Generalization: does it work at a hospital that did not build it? A model that beats radiologists on a benchmark (a fixed test set) has beaten them on one dataset. Deep-learning models for pneumonia on chest x-rays were trained at one hospital system and tested at another. Performance dropped. The models could also tell which hospital an image came from with better than 99.9 percent accuracy. They had learned the scanner as much as the disease (Zech et al., 2018). Radiology also has most of the FDA-authorized devices. Of 903 AI-enabled devices the FDA had authorized by August 2024, 692 (76.6 percent) were radiology devices. Clinical performance studies were reported for only 55.9 percent of all devices, and 2.4 percent were randomized (Windecker et al., 2025). Hundreds of FDA-cleared tools, and so far no measured loss of radiologist jobs. The most widely read explanation of why is Deena Mousa’s essay “AI isn’t replacing radiologists” (October 2025). She also reports that 38 percent of models cleared in 2024 that disclosed their test sites used data from a single hospital. That figure is her reading of the same JAMA Network Open analysis; it is not in the paper’s main text VERIFY: not in the main text of pmid:40305017, checked 2026-09-08; the site counts would be in Supplement 1, which a person should open.4
3. Liability: who signs, and who is blamed when it is wrong? Someone signs the read. Malpractice law judges the physician, not the model. The joint statement of the American, Canadian, European, Australasian, and North American radiology societies describes current tools as helping the radiologist’s interpretation, not replacing it. Tools that act on their own are treated as a separate question needing their own evidence (Brady et al., 2024). Where a device stops and who is liable when a tool is wrong belong to the ethics and regulation chapter. The point here is only that the rules are built to keep a human involved in every decision, and rules change more slowly than benchmarks.
4. Demand: if the task gets cheaper, is more of it ordered? Making a task cheaper does not obviously reduce the number of people doing it. This is the second of the course’s three recurring lessons, efficiency gains get eaten by volume; economists call it the Jevons paradox. In the nineteenth century, more efficient steam engines led to more coal being burned, not less, because cheaper power got used more. Imaging works the same way: cheaper scans get ordered more often. The Neiman Health Policy Institute projects 17% to 27% more imaging in 2055 than in 2023, driven mostly by population growth and ageing, if per-person use stays flat (Christensen, Drake, et al., 2025). Its companion paper projects the radiologist workforce growing 26% to 40% over the same years (Christensen, Parikh, et al., 2025). The workforce projection above assumes demand keeps growing at least as fast as supply. The same paradox appears again in the environmental-impact chapter, if you take it. Efficiency per unit and total use move in opposite directions more often than you would expect.
flowchart TD
accTitle: The four questions, in the order they usually settle an argument
accDescr: Flowchart. Pick one task in the job. Can a machine be measured on it? No, stays with the human. Yes, does it work at a hospital that did not build it? Not shown, stays with the human. Yes, who signs and who is liable when it is wrong? The physician, assists the human. The tool, is demand capped? No, more gets ordered, more of the task, same people. Yes, fewer people do this task.
T[Pick one task in the job] --> Q1{Can a machine be<br/>measured on it?}
Q1 -- No --> H[Stays with the human]
Q1 -- Yes --> Q2{Does it work at a hospital<br/>that did not build it?}
Q2 -- Not shown --> H
Q2 -- Yes --> Q3{Who signs, and who is<br/>liable when it is wrong?}
Q3 -- The physician --> A[Assists the human]
Q3 -- The tool --> Q4{Is demand capped?}
Q4 -- No: more gets ordered --> M[More of the task, same people]
Q4 -- Yes --> R[Fewer people do this task]
Groups working through this usually settle on the first question without being told to. That is worth noticing. Most physician jobs do not break neatly into tasks. The specialties that look most automatable from outside are the ones whose pattern-recognition work is easiest to see, not the ones where it is the largest share of the work.
ImportantA concrete case where the habit gets the wrong answer
Ask your chatbot, “Will AI replace radiologists?” You will get a fluent, balanced answer that mentions image analysis, daily work, and the parts of the job that need a person. Now ask it what fraction of a radiologist’s day is spent interpreting images, with a source. Most models give a number without a source, or a source that does not contain the number. The 2012 stopwatch study above is one paper with fourteen radiologists. A model that has not read it is guessing, and a confident-sounding guess is the whole failure this chapter is about. You will do this properly in the Do block.
9.3.3 The precedent: mammography CAD
The closest earlier example went further than most students expect. Computer-aided detection (CAD) for screening mammography is software that marks suspicious spots on the image for the radiologist. The FDA approved it in 1998. Medicare began paying extra for it in 2002. By the late 2000s it was used on most screening mammograms in the United States, at a cost of over $400 million a year (Lehman et al., 2015). Measured by adoption, it was a success.
Then the outcome data arrived. Across 429,345 mammograms at 43 facilities, adding CAD lowered specificity (the share of healthy women correctly called normal) from 90.2 to 87.2 percent. It raised the biopsy rate by 19.7 percent. It did not significantly change the cancer detection rate (Fenton et al., 2007). A later analysis of 625,625 digital mammograms found no improvement on any measure with CAD. Among radiologists who read both with and without it, sensitivity (the share of cancers found) was lower with CAD (Lehman et al., 2015). From January 2018 Medicare stopped paying separately for CAD. The new mammography codes include CAD “when performed”, so the add-on disappeared into the base payment (CMS Transmittal 3844, 2017). A technology was adopted at scale, generated extra procedures for patients, and then disappeared once the payment stopped.
The finding to remember is smaller and worse. Readers were shown mammograms in which an unusually high share of the CAD marks were wrong. Their sensitivity fell compared with readers who had no computer output at all. It fell most on exactly the cancers the computer had missed (Alberdi et al., 2004). The name for this is automation bias: the habit of trusting a machine’s suggestion more than it deserves. It has been documented in aviation, in decision-support software, and in imaging (Goddard et al., 2011). CAD did not replace the radiologists. It made some of them worse. That is a different and more troubling failure than the one everyone was worried about. The same thing happens with the sepsis alert you will silence in your first month. A tool that is wrong in a systematic way makes the human worse rather than being corrected by them. The defence is the same in both places: find out how often the tool is wrong before you decide how much to trust it.
Adoption and outcomes are separate questions. Objective 3 is that sentence.
9.3.4 The economics the debate leaves out
Everything so far points toward optimism about physician jobs. This section complicates that, because “your job survives” and “the system gets better or cheaper” are different claims.
Within the organization. the critical appraisal chapter follows ambient scribes (software that listens to a visit and drafts the note) through their levels of evidence. The five-system study of 8,581 clinicians found that scribe users spent 13.4 fewer minutes a day in the EHR and 16.0 fewer on documentation. They delivered 0.49 more visits per week. After-hours EHR time did not change significantly (Rotenstein et al., 2026). At UCSF, across 1,565 physicians and about 1.2 million outpatient visits, scribe users billed 1.81 more relative value units per week, a 5.8 percent increase. (RVUs are the units Medicare uses to price physician work.) They saw 0.80 more patients per week and had no rise in claim denials (Holmgren et al., 2026). Time saved became visits and RVUs. That is real, and it is what a hospital’s finance office wants. It is not cost reduction. It is the same recurring lesson, efficiency gains eaten by volume, seen on a billing statement.
Across the system. The Peterson Health Technology Institute takes no vendor money. In March 2025 it concluded that scribes probably reduce burnout but show little evidence of financial return. In April 2026 it concluded that administrative AI reduces work inside organizations while failing to lower system-wide costs, and possibly raising them.5 The pattern it names is “bot wars.” Providers use AI to submit more prior-authorization requests (the insurer’s approval needed before a treatment) and to code visits at higher levels. Payers use AI to process more denials and to “downcode” claims automatically, paying at a lower level than billed. The number of requests and denials rises, and no clinical question is settled any faster. PHTI’s April 2026 report also records a health system where Level 5 visit coding (the highest billing level for an office visit) rose 5 percent after scribes arrived, worth over $1,000 per provider per month. That is the UCSF finding seen from the payer’s side.
One number to refuse to simplify. You will hear that “25 percent of health care spending is administrative.” That sentence blends three studies with three different denominators, that is, three different answers to “percent of what?” Shrank and colleagues estimated total waste at $760 to $935 billion, about 25 percent of spending. Administrative complexity is one of six categories of that waste (Shrank et al., 2019). Himmelstein and colleagues found administration to be 25.3 percent of hospital spending, the highest of eight countries (Himmelstein et al., 2014). Woolhandler and colleagues put administration at $294.3 billion in 1999, 31.0 percent of spending after exclusions, using the broadest definition in the literature (Woolhandler et al., 2003). Three numbers, three denominators. The habit of asking “percent of what?” will outlast all three figures.
9.3.5 The most useful single fact: the BLS contradiction
The Bureau of Labor Statistics (BLS) is the United States government agency that publishes ten-year employment projections for every occupation, in its Occupational Outlook Handbook. Three of those occupations sit side by side in the medical records department (Figure 9.4).
| Occupation | 2025 jobs | Projected change, 2025–35 | BLS’s stated reason |
|---|---|---|---|
| Medical transcriptionists | 42,000 | −4% | Speech recognition and natural language processing let physicians document in real time |
| Medical records specialists | 200,700 | +8% | Aging population, chronic disease, more coding for reimbursement; AI coding tools “may affect” demand |
| Health information technologists and medical registrars | 42,000 | +16% | Growing volume of electronic health information needing analysis and registry maintenance |
What separates them is the same four questions, applied to non-physician work. Transcription is a single task a machine can be tested on. It works anywhere, because speech is speech. Nobody signs a transcript. And its demand is capped by the number of visits. Coding and health-information work fail the first test. The job is judgment about what a record means and what a payer will accept, and the bot wars above are increasing that work, not removing it. The people whose jobs change first are your colleagues, and the harm is not spread evenly. Transcription pays less than the two occupations that are growing, so the displaced workers are the ones with the least money in reserve. The bias and equity chapter has a name for that, harms of allocation, and it applies to jobs as much as to diagnosis.
9.4 Watch or listen
First, the source (1 min 24 s). “Geoff Hinton: On Radiology,” Creative Destruction Lab, YouTube, uploaded 24 November 2016. https://www.youtube.com/watch?v=2HMPRXstSvQ. Watch it before reading anything about it. Notice how specific the claim is, how confident the tone is, and that the whole argument is about image analysis. He is describing a task and forecasting a job. That difference is what this chapter is about.
Then, the other side (watch the first 20 minutes). “AIMI Grand Rounds: Developing Clinically Useful AI for Radiology,” Curtis Langlotz, Stanford AIMI, YouTube, 30 January 2025, 59 min 37 s. https://www.youtube.com/watch?v=WS9p0wFw7uY. Langlotz is the radiologist who wrote, in 2019, that AI will not replace radiologists, but radiologists who use AI will replace those who do not (Langlotz, 2019). Listen for what he says makes a tool clinically useful, as opposed to merely accurate. Listen also for every place where the obstacle he describes is one of the four questions rather than model performance. Twenty minutes is enough for the exercise. The rest is worth hearing when you have time.
If you would rather read than watch, Mousa’s essay in the footnote above covers the same material in about the same time.
9.5 Do
WarningNo patient information, and no classmate’s data
Every prompt in this exercise is about a specialty, an occupation, or a published number. None is about a patient, a rotation, or a person you know. Your institutional chatbot account does not change the no-PHI rule (no protected health information, ever). Free tiers may use what you type to train the next model. The one thing you will paste is a public government web page.
What counts as patient information, and why taking out the name is not enough: the patient information page.
9.5.1 Part A: grade the model (10 minutes)
Step 1. Open your chatbot and ask, substituting your intended specialty (or radiology if you are undecided):
I am a fourth-year medical student choosing a specialty. Will AI replace
<specialty>by 2036? Give me your best forecast and your reasoning.
Step 2. Grade the answer against the four questions. The question for each row is not “did it mention this” but “did it give me something I could check.”
Tool: ____________________ Specialty: ____________________
Factor Mentioned? Gave a checkable number or source? Note
Task decomposition Y / N Y / N ______
Generalization Y / N Y / N ______
Liability/signature Y / N Y / N ______
Demand Y / N Y / N ______
Its forecast in one line: ______________________________________________
What fraction of the job did it say is automatable? ______ Source? Y / NStep 3. Ask one follow-up:
What fraction of a
<specialty>physician’s working time is spent on the task you think is most automatable? Give me the study you are relying on, with authors, journal, and year.
Search PubMed for whatever it gives you. If the study exists and contains the number, note it. If it does not, you have just seen a confident forecast with no evidence behind it. That is the lesson of the 2016 clip at a smaller scale. If you have not opened a chatbot in a year, this exercise is for you. It takes ten minutes, it needs no skill with prompts, and the model failing is the expected result, not your failure.
9.5.2 Part B: your own forecast (15 minutes)
Step 1. Find out how the time is actually spent. The 36 percent figure for radiology came from someone with a stopwatch. Most specialties have such a study; the term to search for is “time and motion study.” Search PubMed with:
"time and motion studies"[MeSH] AND <your specialty>adding physician or resident if the results are mostly about nurses. Read the abstracts of the two most recent studies that observed physicians rather than surveying them. Write down how the day divides. If your specialty has no such study, say so on the sheet. That absence is itself a finding about how well the job is understood. If there is none, use your own sub-internship, labeled honestly as one observer’s estimate.
Step 2. Write three lines.
Specialty: ____________________
1. The most automatable TASK in this specialty (a task, not the job):
__________________________________________________________________
2. What fraction of the working day that task is, and where the number
comes from (study / my own estimate):
______ % source: ____________________________________________
3. The factor most likely to make my estimate wrong, and why:
[ ] task decomposition [ ] generalization [ ] liability [ ] demand
__________________________________________________________________Radiology’s 36 percent is the number to compare yours against. Most students overestimate the fraction for their own field when they first write it down. That is fine. The point of writing it down is to have a number you can check later.
Step 3. Argue against it. Reread your line 2 as the uncle at dinner. What would he say is missing? Then reread it as the radiology resident in the second press clipping. Change the number if either of them convinced you, and say which.
9.5.3 Part C: close a flag yourself (10 minutes)
The module this chapter was built from carried the BLS projections marked “unverified,” because the government site blocked the automated download. You are going to do what the software could not.
Step 1. Open the three BLS pages linked under Table 9.1 in your browser. For each, find the “Job Outlook” section and record:
Occupation Projected change Decade BLS's stated reason
Medical transcriptionists ______ % ____–____ ___________________
Medical records specialists ______ % ____–____ ___________________
Health information technologists ______ % ____–____ ___________________If your numbers differ from Table 9.1, BLS has revised the projections since this chapter was written. Note the new decade, and trust the page over the chapter.
Step 2. In two sentences, explain why the first occupation moves one way and the other two move the other way. Use at least two of the four questions by name. Then one more sentence: who bears the cost of the −4 percent, and what does that group look like compared with the +16 percent one?
9.6 Check
Score yourself against this before moving on. Objective numbers refer to the list at the top of the chapter.
| You have | Meets | Falls short |
|---|---|---|
| The prediction and its outcome (obj. 1) | You can state the 2016 claim, give two primary numbers from 2026 (positions, fill rate, or workforce projection), and say in one sentence why “early” is the wrong diagnosis | “He was wrong” with no number, or “he was just early” |
| A graded chatbot answer (obj. 2) | All four rows filled, and you searched PubMed for the study it cited | Rows ticked without checking anything |
| Your three-line forecast (obj. 2, 6) | A task, not a job; a fraction with a named source or an honest “my estimate”; a factor with a reason | “Documentation, 50%, probably fine” |
| The CAD sentence (obj. 3) | “Adoption and outcomes are separate questions,” with the biopsy-rate finding attached | Only the slogan |
| The system-cost sentence (obj. 4) | You can explain in your own words why 5.8 percent more RVUs is not a cost reduction, and what “bot wars” means | “AI saves money” or “AI wastes money” without a mechanism |
| Three BLS rows (obj. 5) | Numbers and reasons read from the live pages, and a two-sentence explanation that names two factors | Numbers copied from this chapter |
| The “percent of what” habit | You can name the three denominators behind “25 percent administrative” | One number, one study |
9.7 What this changes for you
What does this change about how you’ll practice, and what would you have to see in the next five years to conclude you got this wrong?
Write both halves down. For the first, go back to the dinner table. Your uncle is still holding his phone. What do you say about the 2016 quote, using one number from this chapter and one of the four questions? If you are not going into radiology, answer for your own specialty, with your Part B line 2 as the number. For the second half, name the observation that would make you change your forecast: a match statistic, a workforce projection, a BLS revision, a change in who signs the report. A forecast that nothing could prove wrong is not a forecast. It is a feeling.
Then keep the sheet from Part B. Reread it at the end of residency. Every student who does this exercise has just made a confident forecast about a medical specialty. So did a Nobel laureate, with better information than any of you, and he got the mechanism wrong. The point of writing it down is to be able to check.
9.8 Summary
- The most confident job-elimination prediction in modern medicine has had twice its five-year deadline. Radiology offered more positions in 2026 than ever, filled 98.4 percent of them, and projects a larger workforce in 2055.
- It failed because a task was mistaken for a job, not because the timeline slipped. Four questions, task decomposition, generalization, liability, and demand, tell the two apart.
- Radiologists spend about a third of their time interpreting images. Most physician jobs do not break neatly into tasks.
- Mammography CAD was adopted at scale, raised biopsies without raising cancer detection, made some readers worse, and disappeared once the payment stopped. Adoption and outcomes are separate questions.
- Scribes turn saved time into visits and RVUs. That is real, and it is not cost reduction. Across the system, administrative AI may be raising costs through bot wars.
- Transcription −4 percent, coding +8, health information +16, in the same department. The four questions explain the split, and the harm falls on the lowest-paid of the three.
- “Percent of what?” is a habit worth more than any of the three numbers.
- Your uncle, and your chatbot, are forecasting a job from a task. Now you can say why that is the mistake.
9.9 Go deeper
Papers
- The time-and-motion study behind the 36 percent (Dhanoa et al., 2013), and the editorial that gave the field its slogan (Langlotz, 2019).
- The two CAD studies (Fenton et al., 2007; Lehman et al., 2015) and the automation-bias experiment (Alberdi et al., 2004), read together as one story.
- The generalization paper (Zech et al., 2018), which is short and has a figure you will remember.
- The Neiman workforce papers: supply through 2055 (Christensen, Parikh, et al., 2025), attrition (Christensen et al., 2026), and the pediatric subspecialty (Morales-Tisnés et al., 2025).
- The multi-society statement on buying and monitoring radiology AI (Brady et al., 2024), which is what a department actually does with a cleared tool.
- The scribe economics: the five-system study (Rotenstein et al., 2026) and the UCSF productivity analysis (Holmgren et al., 2026). The five-system study is also in the appraisal chapter.
- The three administrative-cost papers (Himmelstein et al., 2014; Shrank et al., 2019; Woolhandler et al., 2003), for the denominators.
Readings and talks
- Deena Mousa, “AI isn’t replacing radiologists” (October 2025). The four questions in this chapter are a short version of her argument.
- Peterson Health Technology Institute, Administrative AI: Current Use and Potential Impact (April 2026), and KFF’s interview with its director, “The AI Arms Race in Administrative Health Care” (July 2026).
- A history of AI in medicine tells the IBM Watson story, from Jeopardy! to MD Anderson setting it aside in 2017, and covers MYCIN and the AI winters before it. Read it if you want a second example of the same shape before deciding the Hinton story is about one man’s bad call. Also in the foundations chapter.
- AI and machine learning for an oncology service line keeps apart the four groups the economics section needs to keep apart: patients, providers, researchers, and administrators. It also lists eight guiding principles (Badal et al., 2023). Two of them, “have high healthcare value” and “reduce overdiagnosis and overtreatment,” are the CAD story stated as rules. The deck is also in the ethics and regulation chapter; the principles are also in the appraisal chapter.
Alberdi, E., Povykalo, A., Strigini, L., & Ayton, P. (2004). Effects of incorrect computer-aided detection (CAD) output on human decision-making in mammography. Academic Radiology, 11(8), 909–918. https://doi.org/10.1016/j.acra.2004.05.012
Badal, K., Lee, C. M., & Esserman, L. J. (2023). Guiding principles for the responsible development of artificial intelligence tools for healthcare. Communications Medicine, 3(1). https://doi.org/10.1038/s43856-023-00279-9
Brady, A. P., Allen, B., Chong, J., Kotter, E., Kottler, N., Mongan, J., Oakden-Rayner, L., Pinto Dos Santos, D., Tang, A., Wald, C., & Slavotinek, J. (2024). Developing, Purchasing, Implementing and Monitoring AI Tools in Radiology: Practical Considerations. A Multi-Society Statement From the ACR, CAR, ESR, RANZCR & RSNA. Journal of the American College of Radiology : JACR, 21(8), 1292–1310. https://doi.org/10.1016/j.jacr.2023.12.005
Christensen, E. W., Drake, A. R., Parikh, J. R., Rubin, E. M., & Rula, E. Y. (2025). Projected US Imaging Utilization, 2025 to 2055. Journal of the American College of Radiology : JACR, 22(2), 151–158. https://doi.org/10.1016/j.jacr.2024.10.017
Christensen, E. W., Liu, C.-M., Rula, E. Y., & Parikh, J. R. (2026). Attrition of the National Radiologist Workforce: Associations With Radiologist and Practice Characteristics. AJR. American Journal of Roentgenology, 226(1), e2533587. https://doi.org/10.2214/ajr.25.33587
Christensen, E. W., Parikh, J. R., Drake, A. R., Rubin, E. M., & Rula, E. Y. (2025). Projected US Radiologist Supply, 2025 to 2055. Journal of the American College of Radiology : JACR, 22(2), 161–169. https://doi.org/10.1016/j.jacr.2024.10.019
Dhanoa, D., Dhesi, T. S., Burton, K. R., Nicolaou, S., & Liang, T. (2013). The evolving role of the radiologist: The Vancouver workload utilization evaluation study. Journal of the American College of Radiology : JACR, 10(10), 764–769. https://doi.org/10.1016/j.jacr.2013.04.001
Fenton, J. J., Taplin, S. H., Carney, P. A., Abraham, L., Sickles, E. A., D’Orsi, C., Berns, E. A., Cutter, G., Hendrick, R. E., Barlow, W. E., & Elmore, J. G. (2007). Influence of computer-aided detection on performance of screening mammography. The New England Journal of Medicine, 356(14), 1399–1409. https://doi.org/10.1056/nejmoa066099
Goddard, K., Roudsari, A., & Wyatt, J. C. (2011). Automation bias: A systematic review of frequency, effect mediators, and mitigators. Journal of the American Medical Informatics Association : JAMIA, 19(1), 121–127. https://doi.org/10.1136/amiajnl-2011-000089
Himmelstein, D. U., Jun, M., Busse, R., Chevreul, K., Geissler, A., Jeurissen, P., Thomson, S., Vinet, M.-A., & Woolhandler, S. (2014). A comparison of hospital administrative costs in eight nations: US costs exceed all others by far. Health Affairs (Project Hope), 33(9), 1586–1594. https://doi.org/10.1377/hlthaff.2013.1327
Holmgren, A. J., Fenton, C. L., Thombley, R., Soleimani, H., Croci, R., DeMasi, O., Byron, M. E., Murray, S. G., Adler-Milstein, J. R., & Yazdany, J. (2026). Ambient Artificial Intelligence Scribes and Physician Financial Productivity. JAMA Network Open, 9(1), e2553233. https://doi.org/10.1001/jamanetworkopen.2025.53233
Langlotz, C. P. (2019). Will Artificial Intelligence Replace Radiologists? Radiology. Artificial Intelligence, 1(3), e190058. https://doi.org/10.1148/ryai.2019190058
Lehman, C. D., Wellman, R. D., Buist, D. S. M., Kerlikowske, K., Tosteson, A. N. A., & Miglioretti, D. L. and. (2015). Diagnostic Accuracy of Digital Screening Mammography With and Without Computer-Aided Detection. JAMA Internal Medicine, 175(11), 1828–1837. https://doi.org/10.1001/jamainternmed.2015.5231
Morales-Tisnés, T., Pelzl, C. E., Christensen, E. W., Rula, E. Y., & Otero, H. J. (2025). Defining the Pediatric Radiology Workforce in the United States Using Insurance Claims Data. Journal of the American College of Radiology : JACR, 22(11), 1429–1435. https://doi.org/10.1016/j.jacr.2025.07.029
Rotenstein, L. S., Holmgren, A. J., Thombley, R., Sriram, A., Dbouk, R. H., Jost, M., Aizenberg, D., MacDonald, S., Kanaparthy, N., Williams, B., Hsiao, A., Schwamm, L., Murray, S., Byron, M., You, J. G., Centi, A. J., Iannaccone, C., Frits, M., Landman, A. B., … Mishuris, R. G. (2026). Changes in Clinician Time Expenditure and Visit Quantity With Adoption of Artificial Intelligence-Powered Scribes: A Multisite Study. JAMA, 335(16), 1408–1417. https://doi.org/10.1001/jama.2026.2253
Shrank, W. H., Rogstad, T. L., & Parekh, N. (2019). Waste in the US Health Care System: Estimated Costs and Potential for Savings. JAMA, 322(15), 1501–1509. https://doi.org/10.1001/jama.2019.13978
Windecker, D., Baj, G., Shiri, I., Kazaj, P. M., Kaesmacher, J., Gräni, C., & Siontis, G. C. M. (2025). Generalizability of FDA-Approved AI-Enabled Medical Devices for Clinical Use. JAMA Network Open, 8(4), e258052. https://doi.org/10.1001/jamanetworkopen.2025.8052
Woolhandler, S., Campbell, T., & Himmelstein, D. U. (2003). Costs of health care administration in the United States and Canada. The New England Journal of Medicine, 349(8), 768–775. https://doi.org/10.1056/nejmsa022033
Zech, J. R., Badgeley, M. A., Liu, M., Costa, A. B., Titano, J. J., & Oermann, E. K. (2018). Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: A cross-sectional study. PLoS Medicine, 15(11), e1002683. https://doi.org/10.1371/journal.pmed.1002683
NRMP, Results and Data: 2026 Main Residency Match (May 2026), Tables 1A, 3 and 7A, read from the primary PDF on 2026-09-07. The figure shows advanced (PGY-2) diagnostic radiology positions; a further 156 categorical PGY-1 positions were offered in 2026 and 148 filled.↩︎
Doximity, 2026 Physician Compensation Report, covering 2025; https://www.doximity.com/reports/physician-compensation-report/2026. The module this chapter is built from quotes the prior year’s figure, $571,749, up 7.5 percent. The two are different survey years, not competing estimates. Quote one, with its year.↩︎
Steve Lohr, “Your A.I. Radiologist Will Not Be With You Soon,” New York Times, 14 May 2025. The direct quotation is as reported by AuntMinnie, because the original is paywalled. The module this chapter is built from dated the interview May 2026; the correct date is May 2025.↩︎
Deena Mousa, “AI isn’t replacing radiologists,” Understanding AI, 1 October 2025, https://www.understandingai.org/p/ai-isnt-replacing-radiologists. The best single reading if you want the whole argument in one sitting.↩︎
Peterson Health Technology Institute, Adoption of AI in Healthcare Delivery Systems: Early Applications and Impacts (March 2025) and Administrative AI: Current Use and Potential Impact (April 2026), https://phti.org/administrative-ai-current-use-and-potential-impact/. The “bot wars” phrase (Exhibit 4, on prior authorization) and the Level 5 coding example (a 5% rise in Level 5 and 7% in Level 4 established-patient encounters after scribe deployment) are in the report PDF, https://www.phti.org/wp-content/uploads/2026/04/PHTI-Administrative-AI-Current-Use-and-Potential-Impact.pdf.↩︎