A MESSAGE FROM OUR SPONSOR

The frontier has moved. Buying tech and AI is easy. Turning it into measurable performance is not. In our work with leading companies, these transformations deliver ~20% EBITDA uplift—but only when they focus on a few domains and change how the work gets done. In Rewired, we show how capabilities turn technology into repeatable, compounding value.

Learn More

“I think we have fallen in the frame as tech nerds of talking about, We’re gonna build superintelligence and… dot dot dot. It’s going to be great for you. And we have not filled in enough of the dot dot dot.”
—Sam Altman, talking to the folks at Core Memory about where he’s gone wrong in his messaging around superintelligence.

And now for the main event…

A decade ago, Geoffrey Hinton, one of the best-known scientists in artificial intelligence, issued a death sentence for an entire profession. “People,” Hinton said, “should stop training radiologists now.” He gave it five years until deep-learning A.I. systems would be outperforming human radiologists to the point of the latter’s sudden and dramatic obsolescence. Of course, his obituary proved premature. The radiology industry, which had roughly 230,000 jobs in 2014 and 270,000 jobs in 2024, is expected to grow 5 percent through 2034, per the Bureau of Labor Statistics. And the demand for imaging services in the U.S. is only expected to grow as the population ages.

Concerns that technology will vaporize the labor force are as old as work itself. But the realities are historically far more nuanced. Nathan Goldschlag, the director of research at the Economic Innovation Group, reminded me that 40 percent of the U.S. workforce worked on farms in the year 1900. Today, it’s 1.2 percent. Automation, he said, didn’t erase the need for labor; it just reallocated us. “A lot of the discussion now is, This time is different because it’s going to be really fast. This time is different because it’s going to hit knowledge workers. This time is different because it’s A.G.I., so it can do the new tasks,” he said.

As always, no one really knows how this new technology will remake the workforce. And as many economists I’ve spoken to point out, we’ve heard This time is different before—and somehow still have jobs today. For his part, Hinton maintained that he was right directionally; he just got the timing wrong. Soon, he told the Times last year, it’ll be “a combination of A.I. and a radiologist, and it will make radiologists a whole lot more efficient in addition to improving accuracy.”

A MESSAGE FROM OUR SPONSOR

The frontier has moved. Buying tech and AI is easy. Turning it into measurable performance is not. In our work with leading companies, these transformations deliver ~20% EBITDA uplift—but only when they focus on a few domains and change how the work gets done. In Rewired, we show how capabilities turn technology into repeatable, compounding value.

Learn More

Hinton’s amended prediction is proving more accurate than the first: A.I. is helping radiologists meet an ever-increasing demand for imaging. “What we’re trying to do,” said Steve Forthuber, the president and C.E.O. of eastern operations at RadNet, one of the country’s largest providers of diagnostic imaging services, “is use A.I. and other forms of innovation and technology to address this shortage in technologists and radiologists, so that we can create greater capacity and meet all this demand.”

Louis Blankemeier, the co-founder and C.E.O. of Cognita, an A.I. radiology startup that was recently acquired by Radiology Partners, told me that he’s intent on building what he described as the “supervised self-driving of radiology,” in which A.I. systems handle the bulk of the workload and their output is checked by the human radiologists who have spent over a decade training for the job. “Geoffrey Hinton does not understand anything about radiology,” he said. “Most people don’t know how complicated radiology is. Even if you’re an A.I. expert, it’s hard to talk about how A.I. can solve a problem when you don’t understand the problem.” He thinks the field of radiology will be “human plus A.I.” for a long time.

Carl Frey, an economist and A.I. professor at the Oxford Internet Institute, reiterated that the Hintonian discourse about radiology is another example of the facile philosophizing about the A.I.-powered jobspocalypse. It’s a “fairly safe bet to say there will be a lot of automation,” he acknowledged—but, he added, everything from consumer preferences to regulatory intervention will slow it down as wage and productivity gains rise to counterbalance it. “The only thing I feel fairly comfortable with is that we will have more automation, and I feel relatively comfortable that that’s going to be a fairly gradual process,” he said. “We will certainly offset a large part of it, but maybe not all. Or we may use A.I. to create new industries and products, and more than offset it.” It’s something of an open question.

In radiology, machine assistance is not a new phenomenon. Computer-aided detection has been around since the late 1980s. Adoption surged into the early 2000s. But in 2007, a study in The New England Journal of Medicine found that these techniques actually reduced radiologists’ accuracy. “It’s this kind of cautionary tale that is the premise for all of my research, which is that small experiments do not translate to real-world use of a technology,” Dr. Hannah Milch, an associate professor of radiology at UCLA, told me. “My research is to prevent mistakes from the past.”

These days, when Milch sits down to interpret a mammogram, she’s no longer the first one to read it. An A.I. system scans it first, categorizing different levels of risk and flagging certain parts of the image for closer inspection. After seeing what the A.I. comes up with, Milch does her full, standard read, then makes a final decision—which, she told me, “may be the same as what A.I. thought, but very frequently it’s different.” Forthuber described a similar scenario, in which radiologists across RadNet’s 425 imaging centers use A.I. as an initial pair of eyes. The result, he said, has been a measurable increase in the accuracy and efficiency of their mammogram scans. Milch noted that this scenario, in which the A.I. system acts as decision support for radiologists, has become fairly standard practice across the industry.

Behind this transition has been a growing understanding, which Milch also acknowledged, that clinical A.I. models have gotten very good and are getting better, even if we don’t yet have robust evidence to prove it. There’s been only one large-scale randomized controlled trial—conducted in Europe and published in January—that assessed the real-world applicability of A.I. tools for mammography. Though the study supports the notion that these systems can boost radiologists’ accuracy and efficiency, Milch has found that most other data on them doesn’t translate to her practice. “What’s interesting is that [the A.I.] makes egregious errors,” she said. “Its negative predictive value is very good—the bigger issue is the false positives. It doesn’t seem nearly as good as the radiologist in my everyday practice, though the data suggests it’s on par.” To help close the evidence gap, Milch is co-leading the first large-scale randomized controlled test of A.I. in breast cancer screening in the United States—a $16 million, multi-institutional clinical trial that was funded in September 2025.

The obvious question is whether—or, perhaps, when—this kind of work might eventually happen without the radiologist at all. Forthuber, for one, said, “It’s not happening for us at this time.” But, he added: “Could it happen for very routine, absolutely normal scans? Perhaps.” He sees even that eventuality as a positive. If straightforward screenings can be outsourced to A.I., the radiologists can busy themselves with more-complicated cases. “You’re not going to need fewer radiologists,” he added. “You’re still going to need more who can focus on the increased demand.”

Milch also thinks that completely removing the radiologist from the picture is a long way off, but she believes that’s the direction we’re moving in—only it’s not as dire as it sounds. She could see A.I. systems autonomously taking over the “normal” part of the workload—those scans in which nothing is wrong—and predicts a transition period that could well last “many years” where the human radiologists take care of the abnormal scans. “I think there is a shift happening that, from where I sit, is mostly very exciting, because it’s going to make us do better work and help patients more,” she told me. “It will shift the radiologist role, but I think for the better.”

“Radiologists are fine,” Milch added. “We have plenty to do.”

That’s all for today. I’ll see you next week.

Ian

Need help? Review our FAQ page or contact us for assistance. For brand partnerships, email ads@puck.news.

You received this email because you signed up to receive emails from Puck, or as part of your Puck account associated with {{customer.email}}. To stop receiving this newsletter and/or manage all your email preferences, click here.

Puck is published by Heat Media LLC. 107 Greenwich St., New York, NY 10006