In 2016, the AI pioneer Geoffrey Hinton declared that “people should stop training radiologists now” because “it’s just completely obvious that within five years, deep learning is going to do better than radiologists.” He was half right. Today, the FDA has approved more than 1,000 AI radiology tools, some capable of analyzing medical images to detect injuries or diseases with greater accuracy than human specialists. Yet radiologists—human ones—are in more demand than ever. Since 2016, the number of radiologists has risen by 17 percent, the field’s vacancy rates are near all-time highs, and the average salary has increased from about $350,000 to $570,000, making radiology the third-highest-paid medical speciality in the United States.
Many people now fear that AI will make a huge number of careers obsolete. Last year, Anthropic CEO Dario Amodei claimed that AI would soon “wipe out half of all entry-level white-collar jobs.” But the radiologist story suggests that whether AI will replace a given profession is not so straightforward to predict. Answering the following three questions can help you determine how endangered a job really is.
Question 1: Is your job a weak bundle or strong bundle?
According to Luis Garicano, an economist and a co-author of the forthcoming book Messy Jobs, most white-collar jobs combine two very different kinds of work. “Clean” tasks involve predictable problems, objective standards of success, lots of written data, and little interpersonal interaction (think: approving an expense report or updating a spreadsheet). These are the easiest for AI systems to handle.
“Messy” tasks, however, involve dealing with unpredictable situations, meeting subjective measures of success, acting on tacit knowledge, and navigating complex webs of human relationships (think: choosing a new corporate logo, assuaging an upset client, or managing a team). AI isn’t so good at these kinds of tasks, at least not yet. This means that a job’s susceptibility to AI replacement depends, in part, on how easily the clean tasks can be cleaved off from the messy ones.
[From the March 2026 issue: America isn’t ready for what AI will do to jobs]
A trial lawyer has what Garicano and his co-authors call a “strong bundle” job, in which the various responsibilities are so tightly linked that delegating some of them to AI would actually be counterproductive. She might spend most of her time on the relatively clean tasks required to prepare for a given trial—reading relevant case law, studying the facts of the case, drafting an opening argument—and far less time on the messy tasks involved with appearing in court. In theory, much of that trial-prep work could be delegated to a large language model. In practice, doing so would be a huge mistake. During a trial, a lawyer can’t just read from an AI-generated script. She has to cross-examine witnesses, answer questions from the judge, respond to points made by the other counsel, and adjust her strategy based on constantly evolving circumstances. This all requires her to have a thorough understanding of the facts of the case, knowledge of relevant legal precedent, and a familiarity with potential counterarguments. To perform well at the messy part of her job, the lawyer needs to have done much of the clean part herself.
Other jobs are “weak bundles.” A friend of mine who works as a recruiter for a major HR firm used to spend most of his days sifting through résumés. Now AI can easily do that for him. So instead, he spends far more time sourcing potential recruits, talking to hiring managers, interviewing candidates, and negotiating offers. The fact that he’s no longer reading every single résumé doesn’t affect his ability to do the rest of the job. Likewise, delegating the basics of code-writing to AI doesn’t affect an experienced software developer’s ability to do more complex design and engineering tasks.
If AI makes an individual recruiter or software developer more efficient, does that mean that far fewer of them will soon be needed? Not necessarily. What exactly happens to weak-bundle jobs depends on how the rest of the economy responds. In some cases, a job that is largely automated can, somewhat counterintuitively, experience higher levels of employment precisely because it was automated. Whether this occurs depends on the answer to the next question.
Question 2: If what you produce got cheaper, how much more of it would people want?
When the first automobiles were being produced in the 1890s, each car had to be manually built by a large team of workers. Then, in 1913, Henry Ford introduced the assembly line, which could churn out far more cars with far less human labor. The partial automation of car assembly did not cause employment in the industry to collapse; instead, the opposite happened. With fewer workers required to produce each car, factories could make and sell them much more cheaply. Lower prices meant that more consumers could afford to buy a car. Auto manufacturers had to hire many more workers to keep up with the surge in demand. The number of workers in the U.S. automobile industry roughly doubled over a period of 35 years.
A similar story played out with American textile workers following the introduction of the power loom in 1814, bank tellers following the debut of the ATM in 1969, and accountants following the invention of the spreadsheet in 1979. In each case, a new labor-replacing technology seemed poised to kill an existing profession—ATM literally stands for “automated teller machine”—but instead supercharged the growth of that profession, because lower costs led to increased demand. This phenomenon is known as the Jevons paradox, named after William Stanley Jevons, a 19th-century British economist who correctly predicted that the steam engine’s more efficient use of coal would, counterintuitively, cause demand for coal to rise.
Early evidence of a Jevons paradox for AI is everywhere. Job openings for recruiters rose by 30 percent from 2023 to 2025; for software engineers, they’ve doubled. Even as more firms employ AI to handle customer-service requests, the number of call-center workers is rising fast. “It’s not hard to imagine this happening with financial services, with legal services, with health care,” Torsten Slok, the chief economist at the asset-management company Apollo, told me. “As AI makes these services cheaper, people are going to want a lot more of them. And that means employment in those sectors will grow.”
[Lila Shroff: Someone finally wants to hire philosophers]
Efficiency gains don’t always lead to higher demand, however. The share of income that consumers spend on food has fallen by roughly 70 percent since the turn of the 20th century thanks to the mechanization of farming. But about 1 percent of Americans work in farming today compared with about 40 percent back then. People can eat only so much food.
For jobs in which automation doesn’t stimulate more demand, the risk of replacement is higher. But AI’s impact on those jobs will come down to another factor: expertise.
Question 3: Is AI the expert, or are you?
As corporate America began to professionalize during the late 19th and early 20th centuries, companies started employing accounting clerks to keep their financial books in order and inventory clerks to keep track of their wares. Through the 1980s, workers in these roles spent their days on very similar tasks, such as recording transactions, transcribing information, and doing arithmetic. Then came computers capable of automating much of that work. This change affected the two professions in starkly different ways. From 1980 to 2018, the number of inventory clerks nearly tripled, but their average wage fell by 13 percent; the number of accounting clerks, meanwhile, fell by a third, but the ones who remained saw their average wage rise by 40 percent.
According to the MIT economists David Autor and Neil Thompson, the divergence between these two professions boils down to the interaction of technology and expertise. For accounting clerks, computers replaced many of their least expert skills; the hours they had spent recording transactions and performing manual calculations could now be reallocated to more complex tasks, such as explaining why a department had blown through its budget and figuring out sources of discrepancies between a company’s bank account and its books. This turned the accounting clerk’s job from a middle-class one into a smaller, more professionalized one. For inventory clerks, on the other hand, computers replaced their most expert skill set—their encyclopedic knowledge of a warehouse’s physical inventory—leaving them to perform more basic tasks such as scanning items and restocking shelves. This transformed the inventory clerk’s role from a well-paying middle-class profession into a lower-paid job with a far bigger pool of potential workers. In a recent paper, Autor and Thompson find that this basic pattern has held up across more than 300 occupations over the past four decades. “The story is almost never as simple as: We’re in a race with machines and machines will win,” Autor told me. “What matters for a given profession is whether technology enhances a worker’s expertise or commodifies that expertise.”
Applying this framework in the age of AI is not straightforward, in no small part due to the fact that it’s too early to tell just how expert these AI systems will eventually become. According to data from ZipRecruiter, the share of senior-level-job postings in the tech industry has risen considerably over the past year while the share of entry-level-job postings has fallen slightly. But Autor believes that this dynamic could easily change, as AI systems get better and better at engaging in the kind of “expert judgment” that only human experts previously possessed. He pointed to an “electrician’s assistant” tool being piloted by Schneider Electric that allows a normal electrician with only vocational training to troubleshoot the kinds of complex problems that had previously required teams of engineers with graduate degrees. “I think we’re going to begin to see more and more cases like this where AI turns out to be expertise-leveling,” Autor said.
Taken together, these different questions help explain the puzzle of radiology. Radiology is a strong bundle: It combines clean tasks, such as reading and interpreting scans, and messy tasks, such as talking with patients, overseeing imaging exams, explaining results, and making recommendations to clinicians. These responsibilities are highly dependent on one another: Properly interpreting a scan is difficult without intimate knowledge of a patient’s medical history, symptoms, and general health, which can usually be gleaned only by interacting with the patient or their referring physician. To the extent that AI tools have automated part of the job, the radiologists’ remaining tasks require a high level of formal training and specialized knowledge, or expertise. And, as the price of scans has fallen dramatically over the past two decades, clinicians have responded by ordering a whole lot more of them, increasing the demand for even more radiologists, meaning that the Jevons paradox applies.
[Annie Lowrey: How to guess if your job will exist in five years]
This is all a lot easier to evaluate in hindsight, of course. When I tried to apply the framework to my own job as an Atlantic staff writer, the answer was more uncertain. The Jevons paradox doesn’t seem to apply to journalism: Over the past few decades, the cost of a journalism subscription has fallen considerably in inflation-adjusted terms, but readership has plummeted as well. And it’s hard to say which aspects of my work require more expertise: conducting deep research, something AI can do pretty well, or writing a good first draft, which AI—at least for the moment—can’t.
The good news for me is that my job seems to be a strong bundle. It combines clean tasks, such as reading and researching, with messy tasks, such as interviewing experts, discussing ideas with my editor, and writing a good draft. The two parts of the job can’t neatly be separated. I could technically ask AI to scour my call transcripts for key insights, summarize the findings of a paper or book, or come up with questions to ask an expert, but I’ve found I need to do those tasks myself if I want to write and interview well. That should make me hard to automate.
I hope. If there’s one lesson from the history of technology, it is that these changes are hard to predict. Everyone loves to point out that the number of bank tellers rose for decades after the invention of the ATM. But today, the bank-teller profession is indeed dying. It was killed not by the invention that was intended to replace it, but by one that no one expected: the iPhone. When it was invented, no one predicted that this new device would eventually transform how the whole world banked. Some of the most dramatic consequences of the AI revolution are guaranteed to be just as surprising.