# Radiology is a poor proxy for near-term AI job losses

Markdown copy of https://realalpha.app/s/radiology-is-a-poor-proxy-for-near-term-ai-job-losses/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [John Coogan](https://realalpha.app/people/john-coogan/) (Host) · Co-host, TBPN · Co-founded Soylent and Lucy; OpenAI bought his show TBPN in April 2026.
- Host: [Jordi Hays](https://realalpha.app/people/jordi-hays/) · Co-host, TBPN · Founded Branded Native; co-founded fintech Capital, which Rho bought.
- Show: TBPN
- Episode: [Thursday, September 25th](https://realalpha.app/v/QL_GSxeXG5I/), Sep 26, 2025
- Moment: 48:00 to 53:21 of the episode (5 min), playing on https://realalpha.app/s/radiology-is-a-poor-proxy-for-near-term-ai-job-losses/; the episode from there: https://realalpha.app/v/QL_GSxeXG5I/?t=2880

## Quote

> “jobs that look like **repetition of one rote task**, each task being relatively independent, closed, not requiring too much context”
> — [John Coogan](https://realalpha.app/people/john-coogan/) at [52:01](https://realalpha.app/v/QL_GSxeXG5I/?t=3121) · TBPN

## Takeaways

1. Radiologists do more than classify images; their work involves context and communication.
2. Regulation, liability and deployment realities slow adoption.
3. Look for short, repeatable, low-risk digital tasks to predict early automation.

## About this segment

Karpathy's radiology example shows why image-recognition benchmarks don't capture the full job: risk, regulation and patient context matter. He points to short, repeatable, low-risk tasks as likelier early targets.

From “Thursday, September 25th”, Sep 26, 2025. [Watch the full episode](https://realalpha.app/v/QL_GSxeXG5I/)

## Transcript

From the transcript TBPN publishes with the episode.

[48:12](https://realalpha.app/v/QL_GSxeXG5I/?t=2892) Yep. So Andre Karpathy says AI is isn't replacing radiologists. Tyler, do you remember this prediction? Yeah. Was this Hinton? Hinton. Yeah. Yes. Years ago. This was this was, what, ten years ago now or something?

[48:30](https://realalpha.app/v/QL_GSxeXG5I/?t=2910) Yeah. And because this was based off just like image classification. So this was like early on, like CNN's, like Yeah. This was probably mid twenty tens

[48:37](https://realalpha.app/v/QL_GSxeXG5I/?t=2917) Yeah. I assume. So the expectation was that rapid progress in image recognition AI would delete radiology jobs as famously predicted by Jeff Hinton, now almost a decade ago. In reality, radiology is doing great and is growing. So what's going on here? Obviously, image recognition is fantastic. These algorithms are remarkable, and there's been a ton of progress. So Andre Karpathy says there are a lot of, in his opinion, naive predictions out there on the imminent impact of AI on the job market. For example, a year ago, Andre was asked by someone who should know better if he thinks that there will be any software engineer still today, like a year from now. Like, this is how accelerationist this person was.

[49:22](https://realalpha.app/v/QL_GSxeXG5I/?t=2962) And Andre says, spoiler, I think we're gonna make it. This is happening this is happening too broadly. This post goes into detail on why it's not that simple using the example of radiology, and I thought this was really, really interesting, and it obviously applies outside of radiology. So first off, the the benchmarks are nowhere near broad enough to reflect real actual scenarios. So even though you can classify cancer in a radiology environment with a whole bunch of training data very reliably, there's still tons of edge cases where having a human in the loop is advantageous.

[50:00](https://realalpha.app/v/QL_GSxeXG5I/?t=3000) And then just in general, the job is a lot more multifaceted than image recognition. Like if you just think about the job of a radiologist, it's not purely just look at image, detect cancer. Look at image, detect cancer. Look at

[50:15](https://realalpha.app/v/QL_GSxeXG5I/?t=3015) image, Yeah. Imagine if you were getting God forbid, somebody's getting seeing going and getting seeing a traditional office of a radiologist and and they're just watching a television screen that's reporting telling them that something is wrong with them and using a voice model to explain it. Then it just

[50:36](https://realalpha.app/v/QL_GSxeXG5I/?t=3036) is like No. They're just in hinge mode. Swipe right if cancer. Swipe left if it's not. And they're just sitting there swiping all day. Obviously, that's not the job of radiologists. There's so much more. They're doing research on what's changing, integrating new information, talking to different patients about all the different signals, what they could be doing, what they could be putting in their body, how are they feeling, what medicines they're taking. There's a ton of different stuff that they're probably involved in.

[51:02](https://realalpha.app/v/QL_GSxeXG5I/?t=3062) Yeah. Anytime anyone gets any type of lab or report from a doctor Yep. The first thing you wanna do is talk to somebody that's an expert on that, your doctor Yep. And have them explain it to you, the implications, etcetera, good or bad.

[51:15](https://realalpha.app/v/QL_GSxeXG5I/?t=3075) Yep. And then he highlights deployment realities. So regulatory, insurance, liability, diffusion, and institutional inertia. Like, it just takes time to roll these technologies out. But the last one is super interesting to me. He cites Jevan's paradox. So if radiologists are sped up via AI as a tool, a lot more demand shows up. And so if all of a sudden the cost to get a scan or the time that it takes to actually do the scan drops, everyone will say, Well, I want to be scanned all the time. I want to be on top of any sort of development medical that could be detected by a radiologist. And so he concludes by saying, I will say that radiology was, in my opinion, not among the best examples to pick on in 2016.

[52:03](https://realalpha.app/v/QL_GSxeXG5I/?t=3123) It's too multifaceted, too high risk, and too regulated. When looking for jobs that will change a lot due to AI on shorter time scales, I'd look in other places, jobs that look like repetition of one rote task, each task being relatively independent, closed, not requiring too much context, short in time, forgiving the cost of a mistake is low, and, of course, automatable given current and digital capability. Even then, I'd expect to see AI adopted as a tool first where jobs change and refactor, more monitoring and supervising than manual doing, etcetera. Maybe coming up, we'll find a better and broader set of examples of how all of this is playing out across the industry. About six months ago, I was asked I was also asked to vote if we have if we will have less or more software engineers in five years.

[52:56](https://realalpha.app/v/QL_GSxeXG5I/?t=3176) Exercise left for the reader. What do you think he meant, Tyler? What's what's What think? More. Of course. Yes. He he thinks the number of of software engineers will increase, but they will be using AI as a as a tool to lever.