The Job AI Was Supposed to Kill

Summary

---

Podcast Summary: What's Your Problem? – "The Job AI Was Supposed to Kill"

Host: Jacob Goldstein
Guest: Dr. Sourabh Jha, Radiologist, University of Pennsylvania
Date: July 30, 2026
Duration (content): ~00:01:50 – 00:45:42

---

Overview

This episode explores why, despite predictions that artificial intelligence (AI) would quickly automate radiology, human radiologists remain not only essential but in short supply a decade after those bold forecasts. Host Jacob Goldstein and guest Dr. Sourabh Jha discuss the nuanced reality of AI’s impact on radiology, examining both successes and limitations, the role of human expertise, global disparities in technology adoption, and how automation might affect training and professional judgment not just in medicine, but across many fields.

---

Key Discussion Points & Insights

1. The Failed Prediction: AI’s Incomplete Revolution in Radiology

• Geoffrey Hinton's Prediction (02:00):
  Hinton, a pioneer in AI, once declared it "completely obvious" that within five years, AI would outperform radiologists at interpreting medical images.

  "People should stop training radiologists now." — cited by Jacob Goldstein [02:04]

• The Reality, a Decade Later:
  AI has made advances but radiologists are still necessary — and in fact, there’s a shortage of them. The anticipated rapid displacement has simply not materialized.

  "Here we are 10 years later with amazing AI, and we still need radiologists." — Jacob Goldstein [02:46]

• Why Hasn’t AI Automated Radiology?
  Goldstein and Jha dig into where automation falls short—especially when AI’s assistance can be cumbersome rather than simplifying.

---

2. Practical Adoption Hurdles: Why AI Isn’t Yet Replacing Radiologists

• Added Cognitive Burden Rather Than Relief:
  Early AIs would "flag" possible issues on scans, forcing radiologists to check both their own findings and audit the AI’s "suggestions"—often for false alarms (false positives).

  "It slowed me down…" AI felt like "that irritating medical student." — Dr. Sourabh Jha [03:53, 04:00]

• The Mammography Cautionary Tale:
  Early computer-assisted detection in mammography increased unnecessary procedures (biopsies) without improving health outcomes due to over-diagnosing normal findings.

  "It led to more false positives. Women were getting more biopsies...but they were not healthier." — Jacob Goldstein [04:48]

  "Appreciating what normal is… is probably the most challenging task." — Dr. Jha [05:27]

3. The Nuances of Diagnosis and the Value of Expertise

• False Positives vs. False Negatives:
  AI often errs on the side of caution ("can't miss any cancers") but in medicine, the real art is in reassuring patients and not over-investigating.

• AI Successful Use Case: Lung Nodules
  AI helps highlight potentially dangerous spots on lung scans, reducing the workload of finding rare problems amid many benign findings—provided the AI’s accuracy is high and it doesn’t flood the system with false alarms.

  "It’s like trying to find a needle in a haystack. If it’s finding those needles, I don’t need to spend my time looking for them." — Dr. Jha [08:48]

4. Trust in Automation: When Is AI "Good Enough"?

• Threshold for Trust:
  Goldstein and Jha discuss why AI must be nearly perfect before doctors rely on it—to automate, AI must be vastly better than its human counterparts, not just slightly better.

  "If AI is not 99.999% good, it's not useful. If it’s 90%, there’s a 1 in 10 chance it could be wrong, and I have to check everything." — Dr. Jha [15:52]

• Comparison to Driverless Cars:
  Even when "safer on average," any AI faces a higher bar for trust due to social and systemic expectations—much like autonomous vehicles.

  "Driverless cars are already safer than humans, but they're not safer enough for everybody to trust them." — Jacob Goldstein [16:53]

---

5. Limits of AI in Medicine (and Beyond)

• Why AI May Never Fully Replace Radiologists:
  Much disease is nuanced; many "abnormals" overlap with normal. The need for a human "in the loop" is about responsibility and dealing with ambiguous cases that machines can’t resolve.

  "You reverse that—disease is clear-cut...normal has a lot more variability." — Dr. Jha [43:22]

• The Human Body’s "Lack of Imagination":

  "Biology… isn't very imaginative. The way disease announces itself, it's not very imaginative… things will look similar, diseases will look similar, and disease and not disease will look similar." — Dr. Jha [19:10]

---

6. Global Perspectives: AI as a Tool in Resource-Poor Settings

• Success in the Developing World (Everest, India, Africa):
  Radiologists are scarce outside wealthy countries, but portable X-rays plus embedded AI help diagnose TB and other ailments with accuracy comparable to (median) human readers.

  "With AI, its first use cases have been in low and middle income countries...for tuberculosis." — Dr. Jha [25:27]

  "When the AI says no TB, there is no TB." — Dr. Jha [27:28]

  "It's as good as like an okay doctor...which in a place where there are no doctors...is great." — Jacob Goldstein [28:17]

---

7. The Training Dilemma: Automation Bias & Expertise

• Risks for New Physicians:
  Junior doctors may become over-reliant on AI, missing the opportunity to develop their own judgment.

  "If you don't know enough, you don't know when AI is wrong...you start relinquishing more and more things to AI." — Dr. Jha [31:01]

• How to Teach in the AI Era:
  Dr. Jha explains that examinations and training must require independent expertise, so learners can't simply defer to algorithms.

  "If you use AI, that's fine. But your final exams will be without AI, so get used to doing it." — Dr. Jha [32:06]

---

8. Looking Forward: Hopes and Fears

• Not Worried about Job Loss—Worried about Cognitive Decline:

  "I worry about humans being stupid, or more stupid than we actually are...with the cognitive stuff, you can't just do what we do with the biceps with the cognitive work—we have to keep at it." — Dr. Jha [35:10]

• Best Future:
  AI should help humans be smarter, extend capabilities, and decrease cognitive load, helping see things humans might miss, but without eroding core medical knowledge.

  "I really want to use AI to be smarter...to see things that I don't see." — Dr. Jha [35:56]

• On Over-diagnosis:

  "Healthy people getting a full body scan? Don't bother...for every success story there are about 99 where they've been just taken down a rabbit hole." — Dr. Jha [43:58]

---

Memorable Quotes & Moments

• On the Persistence of Human Expertise:

  "You need somebody accountable...if every time AI says something I have to scratch my head and say, are you right or wrong, that’s not quite automation." — Dr. Jha [15:52]

• On Automation Bias:

  "If you're not [an expert]...AI can make you worse." — Dr. Jha [29:01]

• On Literary Parallels:

  "Do you think AI will replace writers or radiologists? First it'll replace mediocre writers before it replaces mediocre radiologists." — Dr. Jha [40:56]

---

Notable Timestamps

• [02:00] — Hinton’s original prediction
• [04:00] — AI slows down radiologists, acts like a “medical student”
• [05:27] — The art of distinguishing "normal"
• [08:48] — Lung nodules: where AI can help
• [15:52] — Why AI must be near-perfect to be trusted
• [19:10] — Limits of what imaging and AI can decipher
• [25:27] — AI’s use in tuberculosis detection in India, Africa
• [27:28] — "When AI says no TB, there is no TB"
• [29:01] — Automation bias risk for trainees
• [35:10] — Fear of human cognitive decline, not job loss
• [43:22] — Why “normal” is harder to define than disease
• [43:58] — Overdiagnosis from unnecessary full body scans

---

Tone & Style

Conversational, deeply explanatory, and laced with dry wit—both Goldstein and Jha use clear analogies (“irritating medical student,” driverless cars), humility, and a scientist’s skepticism to highlight both promise and pitfalls. Dr. Jha brings global perspective and a healthy resistance to hype, ultimately advocating for careful, expert-driven integration of AI into domains where lives are at stake.

---

Conclusion

This episode dissects the gap between AI hype and clinical reality, revealing why radiology—and, by extension, many professional fields—remain stubbornly human despite rapid technological progress. The balance between trust, expertise, and utility is delicate; the best future is one where AI augments, rather than substitutes, rigorous human judgment.

Transcript

[Podcast Host]

This is an iHeart podcast, Guaranteed Human

[Sponsor Announcer]

Today's episode is brought to you by ChatGPT for Business. As a listener of this podcast, you're looking for ways to help teams move faster, make sharper decisions, and turn scattered context into work they can use. ChatGPT for Business can help. ChatGPT for Business gives teams a shared workspace with admin controls, permissions, and access to work and codecs. In ChatGPT, this means your business can move from question to answer and code to rollout quicker. Join over 10 million business and enterprise users worldwide already using ChatGPT for work. Download the ChatGPT desktop app or contact Sales to learn more.

[Jacob Goldstein]

Did you ever notice how you spend

[Gabriel Hunter]

hours shopping online only to pause at

[Jacob Goldstein]

checkout because you wonder if you trust

[Gabriel Hunter]

it enough to hit buy?

[Jacob Goldstein]

Now Agentic Commerce is testing that moment more than ever.

[Gabriel Hunter]

That's where PayPal comes in, with 25 years of checkouts, 400 million consumer accounts

[Jacob Goldstein]

globally, and the benefit of fraud protection. So no matter where a purchase starts,

[Gabriel Hunter]

it ends with trust. Built for payments, growth and agentic PayPal open built for all business. Visit PayPalOpen.com when you think about discovering small brands, what store pops into your mind?

[Jacob Goldstein]

Well, it should be Walmart.

[Gabriel Hunter]

Seriously, Walmart has thousands of small brands and they're all in one place. Just go online or in store, discover and shop. It could not be easier. Every one of these brands has a real story and real people behind it. They're true American success stories. And you can find them all at Walmart. Discover thousands of small [e-mail] today.

[Dr. Sourabh Jha]

Pushkin.

[Jacob Goldstein]

Ten years ago, Geoffrey Hinton, the computer scientist who's one of the founders of

[Gabriel Hunter]

modern AI, came out with a very strong take.

[Jacob Goldstein]

He said, and this is a quote, people should stop training radiologists now. Radiologists, as you probably know, are doctors who, among other things, read medical scans, X rays, CT scans, MRIs.

[Gabriel Hunter]

And Hinton said it was just completely obvious, again, his words, completely obvious that

[Jacob Goldstein]

within 5 years AI would do a

[Gabriel Hunter]

better job than radiologists at doing that, at reading scans.

[Jacob Goldstein]

And this makes sense. Radiology is largely pattern matching. Does this scan match the pattern of scans that show disease, or the pattern of scans that show a healthy patient? And pattern matching is basically what AI does. And yet here we are 10 years

[Gabriel Hunter]

later with amazing AI, and we still need radiologists.

[Jacob Goldstein]

In fact, there's a radiologist shortage.

[Gabriel Hunter]

So the question is, what happened?

[Jacob Goldstein]

Why do we still need so many radiologists?

[Gabriel Hunter]

And more broadly, what does the story

[Jacob Goldstein]

of AI and radiology tell us about the Future of AI and work for

[Gabriel Hunter]

the rest of us, those of us who are not radiologists.

[Jacob Goldstein]

I'm Jacob Goldstein and this is what's yous Problem?

[Gabriel Hunter]

My guest today is Sourabh Jha. He's a radiologist at the University of Pennsylvania.

[Jacob Goldstein]

He's written about his own use of AI and he's also traveled to India and Nepal and, and has seen the very different ways that AI is being

[Gabriel Hunter]

used for radiology in the developing world.

[Jacob Goldstein]

To start, Sohrab and I talked about why radiologists have adopted AI more slowly

[Gabriel Hunter]

than some people expected they would.

[Jacob Goldstein]

What was it like for you using AI five years ago? Like when you tried to use AI, you know, specifically, sort of what happened?

[Dr. Sourabh Jha]

So to give you an example, when I was using AI for a chest X ray, I found that it slowed me down, quite honestly, because what it would do is it would pick up stuff and highlight it. And now I had to do two things. Firstly, make my own assessment and then make the assessment whether AI was right or wrong. And whereas before I didn't have to do that. So AI was kind of like that irritating medical student. I didn't mean to say that. They're all very nice.

[Jacob Goldstein]

Go on, like, what was it like? Be like, hey, what about this? Should we worry about this?

[Dr. Sourabh Jha]

Yeah. And for most parts, it wasn't really finding things that I wasn't finding, but it was finding things that weren't real. So it increased my cognitive burden.

[Jacob Goldstein]

There is this interesting story from kind of before the modern era of AI. There was computer assisted mammography. Right. There was an earlier version of this and as I understand it, in that earlier version, in fact it was unhelpful. Right. It led to more false positives. Women were getting more biopsies, more follow up tests because they maybe had cancer and they didn't. And in the end it was not improving outcomes. Right. So I understand that that's an old story, but like, that points to the risk of AI. It's not always good to have more flags raised.

[Dr. Sourabh Jha]

Yeah, I'm glad you raised that because that does address two very important things. You know, we think about physicians as the sort of people that spot diseases and as somebody that I trained a fairly long time ago, but not that long time ago. The easy bit was finding disease. The easy bit was actually finding people that were very ill. The difficult bit was telling somebody that you're not ill, that you're actually normal. So appreciating what normal is, appreciating that you don't need to come into the hospital you don't need to be investigated left, right and center. You don't need to be on a bunch of medication is probably the most challenging task.

[Jacob Goldstein]

Say more about that. Why do you say that?

[Dr. Sourabh Jha]

Because when disease looks at you, it stares at you straight in the face.

[Jacob Goldstein]

Yeah.

[Dr. Sourabh Jha]

But for many reasons, when you're at that border zone between disease and not disease, it's very fussy, it gets very foggy. And it's at that point that you have to really instill all your expertise to make sure that it's not, that you're not over calling. So the false positive part is a very big part. And what happened in the past was that the mammography, the computer aided detection of findings was fine tuned. You know, it's like a thermostat. You can fine tune it to some extent the way you want. And they said, we can't miss a single case of cancer, make that mistake. It's okay to over call cancer, but we can't have a false negative because that would be disastrous. And so the, the radiologist who had the computer aided now became like a monkey on the shoulder. It kept saying, positive, positive, positive. You know, you know the story of the boy who cried wolf, right? Yeah. So when there truly was a cancer, people, people were like, yeah, are you just being AI or is this truly cancer? So it's not very helpful if it has a lot of false positives. Right.

[Jacob Goldstein]

And in fact, there was a study published in the New England Journal of Medicine. Right. This is not just anecdotal. Like that early system did not help. Like women had more biopsies, more fear of having cancer, but they didn't, they were not healthier. Like the computer was doing more harm than good at that point. Whatever it was 15 years ago, I

[Dr. Sourabh Jha]

would say that things have improved where. And trust is a very important thing. Trust is a major thing. So as AI improves, my trust of using AI increases. And then at some point I just kind of relax and I'm not so uptight with AI. I'm more like, yeah, relinquishing some stuff. So we're getting to that point where we can basically start trusting, not fully trust, but start trusting AI and start asking ourselves, start saying to ourselves, maybe AI has a point.

[Jacob Goldstein]

And so specifically, how does that work in your practice of medicine? What are specific instances where AI is helpful?

[Dr. Sourabh Jha]

Now, let me give you one very big example. A lot of people have lung nodules, which is little dots in their lungs of varying sizes, most of which, the vast majority of which are not a problem. So if somebody says to you, you have a lung nodule, I wouldn't lose sleep at all. But some of which can become cancer, rather, some of which are precursor to cancer. So a big example is AI flagging the lung nodules. Right. So what that does is it reduces the work on my visual field. Like, I don't need to. It's like trying to find a needle in a haystack. If it's finding those needles in a haystack, I don't need to spend my time looking for those needles.

[Jacob Goldstein]

And it needs to find substantially all of them and not find substantially more than there really are.

[Dr. Sourabh Jha]

Right, yeah. And also it also needs to tell me, based on the shape, the size and other aspects, which can become cancer, which can't.

[Jacob Goldstein]

This, presumably, is the harder diagnostic problem

[Dr. Sourabh Jha]

to solve if we're willing to accept a few errors. It's not that difficult a problem to solve because if we can ratchet the thermostat to tell us that the vast majority of nodules, which are not malignant. You say they're not malignant, we're in good shape.

[Jacob Goldstein]

Yes, Well, I mean, there's always going to be some number of errors. And the nature of eliminating errors is the more you eliminate, the harder it is to eliminate the next one, the marginal one.

[Dr. Sourabh Jha]

Exactly, exactly.

[Jacob Goldstein]

And so presumably the trade off with AI is it needs to, on net, make you catch more without taking so much extra time and have so much extra intervention.

[Dr. Sourabh Jha]

Yeah.

[Jacob Goldstein]

So that the benefit exceeds the cost.

[Dr. Sourabh Jha]

Right, yeah, that is. That is pretty much it.

[Jacob Goldstein]

And are we there now with. With. With lung nodules?

[Dr. Sourabh Jha]

I think we can be. I think that we have the algorithms that can tell us when there's a lung nodule and when it's a problem without really compromising on the false positives too much.

[Jacob Goldstein]

When you say we can be, does that mean it's not quite there yet? Where is this now in your practice? Where is it more generally? And then presumably there's this sort of sociology piece of it, for lack of a better word.

[T-Mobile Advertiser]

Right.

[Jacob Goldstein]

There's does the machine actually work? And then there's, do doctors actually use it? And the first is a precondition. The first is necessary, but plainly not sufficient for the second.

[Dr. Sourabh Jha]

Right, right. And so let's answer your first question about the where are we now? So we are still in the relics of the past, which is that we are still obsessed with making sure AI detects correctly and not misses. So we're still in where AI is pointing stuff out by circles and things like that, as opposed to not just pointing it out, but putting it in the report.

[Jacob Goldstein]

I don't really, I mean, I superficially understand the difference between those two, but why is that an important distinction?

[Dr. Sourabh Jha]

Because if you have 10 nodules that you flag and you don't put it in the report. Yeah. Then radiologists will have to put it in the report.

[Jacob Goldstein]

Uh huh. And the 10 nodules that it flags, it's saying like here are nodules, but they're fine. Yeah. And then making the radiologist do that, just drudgery of typing, basically. Yeah, exactly, that, that one. I mean, maybe it's not easy to solve for legal or bureaucratic reasons. Sounds easy to solve that part of it.

[Dr. Sourabh Jha]

It should be difficult. It should be difficult. But you know, in regards to your second question about trust. So trust is something that obviously takes time, but you know, radiologists aren't the sort of people that have shied away from technology. Yeah, we embrace technology. That's how we've grown over the years.

[Jacob Goldstein]

I mean it only exists because of technology. Like presumably you couldn't have had radiologists before you had X rays which are what, 100 and not, not 200 years old? 100. 100 plus years old. Not that old.

[Dr. Sourabh Jha]

A little north. 125. So if we haven't embraced it, it's probably because it hasn't moved the needle for us. And if it moves the needle then it would be self evident. So big example we have from our own technological development was the time and place where, you know, medical imaging, they were like films that you would put up in the view box, like kind of thing. And then it became digitized. So you no longer did that picture archiving system.

[Jacob Goldstein]

Sure.

[Dr. Sourabh Jha]

And what we call packs when it became that it was adopted left, right and center and nobody was complaining about it and nobody was asking for evidence. It was so self evident.

[Jacob Goldstein]

Yes. Although, I mean if you extrapolate that out, it doesn't have the risk of taking your job in the way that AI theoretically might for radiologists at some point. Right. That seems like a potentially meaningful difference.

[Dr. Sourabh Jha]

It is. But I don't think that anybody is saying no to AI because they are afraid of losing their job.

[Jacob Goldstein]

Why are they saying no?

[Dr. Sourabh Jha]

They're saying no because it's not currently being helpful. So if you're going to be helpful, you have to either make radiologists do less or do things faster.

[Jacob Goldstein]

And is there any domain yet where AI is helping radiologists to either do less or do things faster?

[Dr. Sourabh Jha]

As it stands right now, no.

[Jacob Goldstein]

Yeah, interesting. This is kind of a question for the end. But I can't wait because I'm so interested. Do you feel like there is some broader lesson? It's extremely interesting to me that this domain that seems like it should be so good for AI, it's still not helping. Can you make any inference that goes beyond medicine based on that fact?

[Dr. Sourabh Jha]

Let's say historically, any technology that has come through that is automated, it hasn't required the person overseeing the automation to kind of share the process that leads to the automation, the scientific calculator and the log tables. There's a point in time people were using log tables at a point in time and they stopped. But when they stopped, they weren't checking the calculator's output.

[Jacob Goldstein]

Right.

[Dr. Sourabh Jha]

They weren't like, you know, okay, oh, let's just make sure that that six figure calculation it did was correct. And that's what you need to do in radiology, like if it's going to be AI propel through. But if every time AI is saying something I have to scratch my head and say, are you right or wrong? Then that's not quite automation. And so if AI is not 99.999% good, it's not useful. It cannot be 90%. But if it's 90% and there's 1 in 10 chance that it could be wrong, I don't know what that 1 in 10 is going to come from. I still have to treat every single of its output as potential 1 in 10, which doesn't help me.

[Jacob Goldstein]

You could imagine some world that is not our world, that is purely governed by reason, where the AI just has to be better than the radiologist. It doesn't have to be 99.99%.

[Legal Disclaimer Speaker]

Right.

[Jacob Goldstein]

Because, you know, radiologists are imperfect, like all human beings. Right. And like all machines, like all probabilistic machines, but we don't live in that world. Right. For radiologists to trust AI, the AI has to be way, way better than radiologists. To me, the analogy there is driverless cars. Driverless cars are plainly already safer than humans, but they're not safer enough for everybody to trust them everywhere. Does that seem relevant in radiology too?

[Dr. Sourabh Jha]

So, you know, when you say that AI should be better than radiologists in terms of diagnosis, better is not something that can be easily measured. Like you would measure faster. If somebody's faster, they're faster and there's a timing and they're faster. In radiology, somebody can be better, but better at a certain thing and worse at another thing.

[Jacob Goldstein]

I mean, let me give a naive answer and you can tell me why. Why it's not why it's wrong. Fewer false positives and fewer false negatives.

[Dr. Sourabh Jha]

Oh, I mean, if you can do that, great.

[Jacob Goldstein]

Like, that seems better to me.

[Dr. Sourabh Jha]

But yeah, yeah, that is definitely better. Okay. Fewer false positives and fewer false negatives. But I would also like to see if it can do that, because disease is a spectrum. You know, disease is not just a monolith. Certain things are quite so obviously diseased and so obviously not diseased. So where is it performing? At the entire range of the spectrum. And if it is, then there's no

[Jacob Goldstein]

need for the radiologist as you're talking about that. I get the sense that you feel like that is some distant world that may never happen. Is that a correct inference?

[Dr. Sourabh Jha]

It could happen, but there are a few things to bear in mind. First is I think there's this idea that AI has these superhuman skills, which somehow humans can't figure something out. And AI will figure it out.

[Jacob Goldstein]

In some domains this appears to be the case.

[Dr. Sourabh Jha]

Some domains, yes.

[Jacob Goldstein]

This is not the structure of proteins, notably.

[Dr. Sourabh Jha]

Sure. But the biggest problem, and I think this probably recurs quite consistently, is human biology, human pathology. It's not very imaginative. Let's just take something very simple. How disease presents itself, how disease announces itself, so called clinical symptoms. It's not very imaginative. Like if you have shoulder pain, for instance, Shoulder pain could be from muscle problems in your shoulder. Right. It could also be because one of your organs in your abdomen has just burst. And it can also be a very uncommon manifestation of a heart attack. So there's something about the body that just doesn't really have much imagination. So you're also saying shoulder pain, they're very, very different things. Now, of course, that's just a symptom. Imaging is supposed to clarify things even further. And it does to a large extent, but it also fails. So at some point, biology has its own limitation. Biology, in the way it announces itself, has a certain limitation. And no matter how much you aim to try and decipher that things will look similar, diseases will look similar, and disease and not disease will look similar. Yes. So the reason, many reasons, why you wouldn't give it all to AI is because you have to have the second check. And even if that second check isn't going to make much difference, the mere fact of having the second check means that you don't have autonomy in a domain where things are less than 100% perfect. And the human in the loop is for multiple things, not just simply to make the process more accurate, but to have somebody accountable.

[Gabriel Hunter]

We'll be back in just a minute.

[Sponsor Announcer]

Quick one before you jump back in, you're listening for ways to help teams move faster, make sharper decisions, and turn scattered context into work they can use ChatGPT for business can help ChatGPT for business gives teams a shared workspace with admin controls, permissions, and access to work and codecs. In ChatGPT, this means your business can move from question to answer and code to rollout quicker. Join over 10 million business and enterprise users worldwide already using ChatGPT for work. Download the ChatGPT desktop app or contact Sales to learn more.

[Gabriel Hunter]

You probably think of Walmart as the

[Jacob Goldstein]

store that carries just about everything under

[Gabriel Hunter]

the sun and maybe not as the place to discover small brands. Well, two things can be true at the same time because Walmart is home to thousands of small brands founded by

[Jacob Goldstein]

people who had an idea, took a

[Gabriel Hunter]

chance and built something of their own. Behind everyone is a real story and a lot of hard work. So why not take a little time to recognize all the people building small brands across the country and support everything they're creating? Walmart is proud to give those brands a place on its shelves and online. It's never been easier to find and support small brands. So take a look at Walmart. You might come across something new, something unexpected, or a brand whose story speaks to you. You never know what you're missing until you look Discover thousands of small [e-mail]

[T-Mobile Advertiser]

this summer there are 250 reasons to join T Mobile Girls.

[Podcast Host]

Trip booked and switching to T Mobile help make it happen with experience beyond

[T-Mobile Advertiser]

switch and get $100 back plus $650 in value in built in plan benefits. That's $750 in your first year guaranteed.

[Podcast Host]

And honestly that hundred dollars back I'm putting it towards the zipline now.

[T-Mobile Advertiser]

That's how you do summer.

[Podcast Host]

Exactly. Best switch I ever made.

[T-Mobile Advertiser]

This is just one reason to join T Mobile this summer. There are 250 reasons. Find yours

[Legal Disclaimer Speaker]

Savings based on the value of benefits included with experience beyond like entertainment and one year of AAA Classic and DashPass on us. Benefits may require activation CPlan. For details get $100 via virtual prepaid MasterCard with eligible ported, no cash access and expires in six months issued by Sunrise Banks and a member FDIC.

[Jacob Goldstein]

Allow eight weeks tell me about doing X rays at Everest Basecamp.

[Dr. Sourabh Jha]

Yeah, so that is one of big big use case for AI. A few years ago we took an X ray machine which is a portal X ray machine with AI capabilities to the Everest Clinic in Basecamp, where they didn't have X rays before that and they certainly don't have radiologists. One of the big questions there is, does the person have what's known as high altitude pulmonary edema? It's significant because in a high altitude pulmonary edema you need to get the person down. That's the only treatment. It's not just enough giving them medication. So what AI has been able to do in that realm is to give assist to people who aren't radiologists. And that's a big use case for AI in multiple domains where you don't have radiologists and sometimes you just don't have physicians.

[Jacob Goldstein]

So I read that you did a sabbatical not too long ago at a university in India. And I'm curious more generally whether during that sabbatical or otherwise you saw firsthand either people using AI in settings where there are not many radiologists or you saw potential for that use.

[Dr. Sourabh Jha]

So India is a very interesting country because normally when you talk about technology, there's a drift that happens between high income countries to low middle income countries. With AI, it's kind of the opposite, where its first use cases really have been in low middle income countries and fighting things like tuberculosis. So tuberculosis is still very rampant in Africa, in South Asia, such as India. And there's only so far you can go with detecting tuberculosis using the stethoscope or looking at the sputum. To get further, you have to do a chest X ray, but it's possibly enough to do a chest X ray. That chest X ray needs to be red and read there and then. So imagine you're going around a village in India with a portable chest X ray mounted in a bus or something and you take an image and then you send it back to the city and by the time it's read three weeks later you're like, oh, this person is positive. And now you can't find this person. Now this person, you know, has TB and may very well be infecting other people, but you can't find this person because it's three weeks later. Imagine if you can give the answer there and then, and that's what AI is doing. So AI embedded in the chest X ray is giving the answer there and then and it's making a huge impact.

[Jacob Goldstein]

And so is that happening right now like AI, like places where there are no radiologists? People are using AI to diagnose TB or rule out tb.

[Dr. Sourabh Jha]

It's been happening for the last, I would say at Least for the last eight years.

[Jacob Goldstein]

And how good is it? How good is it? How good is the AI?

[Dr. Sourabh Jha]

Excellent, excellent. Answering the question with high negative predictive values. So when the AI says no tb, there is no tb.

[Jacob Goldstein]

So there are no false negatives. They've turned the dial way in that direction. How is the false positive?

[Dr. Sourabh Jha]

Yeah, so there's false positives. It's not the only thing. The next thing is they have to have a blood test to see if they have TB or not.

[Jacob Goldstein]

How does the false positive rate compare to the false positive rate of the median radiologist?

[Dr. Sourabh Jha]

So I would say that if you have a very high end radiologist, like

[Jacob Goldstein]

a huge expert, that's why I gave you median. Median.

[Dr. Sourabh Jha]

So it's probably lower than median, but it's not as good as the expert or it's just below the expert.

[Jacob Goldstein]

So it's pretty good. It's as good as like an okay doctor.

[Dr. Sourabh Jha]

Yeah, it is.

[Jacob Goldstein]

Okay, okay. Which in a place where there are no doctors or certainly not enough to read the X rays is great.

[Dr. Sourabh Jha]

Yeah. But you know, if you ask any physician, are you above average? Most of them will say they're above average.

[Jacob Goldstein]

It's a problem. At some level it's a problem.

[Dr. Sourabh Jha]

Right.

[Jacob Goldstein]

The Lake Wobegon effect in medicine. So you're affiliated with the University of Pennsylvania. I assume that means you teach, you work with residents, med students, is that right?

[Dr. Sourabh Jha]

That's correct.

[Jacob Goldstein]

So I'm curious how younger physicians, physicians in training, how their relationship to AI is different, or is it different than yours?

[Dr. Sourabh Jha]

So that's a very good question. And I think that gets down to the training that is required in the era of AI. And you'll hear all sorts of things about AI and radiologists, AI and physicians, et cetera, et cetera. But a theme that seems to be very consistent is that if you're a real expert, like if you know your stuff, if you know medicine inside out, if you know radiology inside out, AI can definitely make you better. But if you're not, if you're starting off, you're nervous. I wouldn't use the word mediocre, but I'd say inexperienced. There is a big chance AI can make you worse. And that's because of a phenomenon called automation bias, where if you don't know enough, there's a thing about knowledge which is that there's a certain point, I think it's the Dunnar Kruger or Donna Krueger, I can't remember the name.

[Jacob Goldstein]

Dunning Kruger.

[Dr. Sourabh Jha]

Dunning Kruger, yeah.

[Jacob Goldstein]

People who aren't very good Overestimate their abilities.

[Dr. Sourabh Jha]

That's the basic that is that right? Yeah. But you know, as you go along that curve, there comes a point where you know so much that you start knowing when you don't know. Like you're actually familiar with your ignorance.

[Legal Disclaimer Speaker]

Right?

[Dr. Sourabh Jha]

Yeah. And at that point, at that level of expertise, you can use AI intelligently because you won't use AI to answer questions that you already know, noting that AI can be wrong in this. Rather you know so much that you now you know what you don't know and you can use AI to fill in that gap. What happens on the other extreme, the inexperience extreme, is that you don't have that knowledge, so you don't know when AI is wrong and you start relinquishing more and more things to AI because you never had the point where there was no AI, where you were making those decisions on yourself and you could end up never getting to the point where you could question AI.

[Jacob Goldstein]

So this is the trap. Residents, doctors, training to be radiologists could

[Dr. Sourabh Jha]

fall into, could fall into, could fall into. And I don't think we've kind of figured out quite what to do with that.

[Jacob Goldstein]

Uh huh. That is an interesting problem that seems to generalize. Right. You could imagine that trap in all sorts of fields where, I mean, even, you know, people worry, that kind of entry level jobs that senior people view as drudgery, but that are in fact the way you learn the field. You kind of get this sort of intuition by doing a thing 10,000 times. Those, to the extent they're being automated, seem to present that kind of risk in many fields potentially. But so what do you do about it? You are responsible for training radiologists. How do you help them not fall into this trap?

[Dr. Sourabh Jha]

So you do what? Both my kids are undergrads and they're told very specifically not to use ChatGPT. They said if we catch you, you're not going to get your grades. And better than that is just to examine them in scenarios where they can't access it. So I think we're in a trial and error situation. We're in an experimental area. We haven't quite got any precedence for this in the exact manner, but in order to leverage AI, you've got to be in a scenario where you're a complete physician yourself. And you can't be a complete physician if you start using AI early or too early. And so that's going to require in the form of training to force people or at least say to people. Well, if you use AI, that's fine. But your final exams will be without AI, so get used to doing it. Without AI

[Jacob Goldstein]

Yeah, I mean, I guess it would seem that in radiology it's easier to give people an exam that is like what you actually do as a radiologist than it is in other domains of medicine.

[Dr. Sourabh Jha]

Absolutely. And in fact, the American Board of Radiology, the one that conducts the exam, years ago, when I was doing the exam, it was an oral exam, one on one with the examiner, no AI, no nothing, then just scans.

[Jacob Goldstein]

They just show you a scan and say, what is it?

[Dr. Sourabh Jha]

Just show scans. And on the other side of the scan is somebody with an expressionless face. You don't know when you're right, you don't know when you're wrong. The expression doesn't change. And then they went virtual and they gave questions and multiple choice questions. You still couldn't use AI, but you had no person overlooking that now they're going back to the old system. And I think that's a good thing.

[Jacob Goldstein]

So that people can't cheat with AI,

[Dr. Sourabh Jha]

so that people just learn to talk to people. I mean, that's the most important thing. Most important thing is that you're, you know, you're able to look somebody in the eye and say, hey, you know, I'm reading this scan. Here's what I think it is. And that's the best way to make sure that people are performing like humans. I mean, humans have to be humans. I mean, to be in the loop, you have to be an expert. How could you be in the loop if you're not an expert?

[Jacob Goldstein]

So I want to talk about the future for a minute and maybe there's the future that you're worried about and the future that you hope comes to pass in radiology with respect to AI, do you worry? Is there a future you worry about in radiology?

[Dr. Sourabh Jha]

I'm not worried about my job being taken. I think I don't lose sleep about that.

[Jacob Goldstein]

Would you worry if, like your kid who's in college wanted to be a radiologist?

[Dr. Sourabh Jha]

I'd say that you should look at 10 years and be willing to pivot. So the thing is what I worry about, I worry about humans being stupid, or more stupid than we actually are at the moment, stupider. And I think that it's important that somehow we don't do much manual work and we go to the gym every now and then. But with the cognitive stuff, you can't just do that. Do what we do with the biceps with the cognitive work, we have to keep at it. Otherwise it really will kind of disappear. So I do worry about that cognitive

[Jacob Goldstein]

disappearance, that people will stop thinking and get dumber because we have stopped thinking.

[Dr. Sourabh Jha]

Exactly. I really want to use AI to be smarter and to be able to extend my cognitive abilities and to see things that I don't see. And that's fundamentally it, to see things that I don't see and to be able to process large amounts of information that I can't process because I don't have the time, I don't have the bandwidth and I don't have the computational power. And so I'd like to see AI doing that and I'd like to be able to do that. You know, our knowledge is vast. A lot of things have changed over the years. We used to memorize the bones of the hand. Right, the bones of the hand. There was a nice mnemonic and they're all this very kind of your cheesy mnemonic. So I can't say it. It's not safe for work.

[Jacob Goldstein]

I want you to say it now.

[Dr. Sourabh Jha]

It involves a lot of names and other acts that. Scaphoid, lunate, triquetry, pisiform. You probably don't need to learn that. You probably don't need to memorize all of that. So there's definitely some change in what. How we do things, in sort of. We can't be doing what we were doing in the 70s, 80s and 90s. Now having said that, there's. There's a core. There's a core of being what a physician is. There's a core level of being totality of knowledge that's important to have.

[Jacob Goldstein]

We'll be back in a minute with the lightning round.

[Sponsor Announcer]

Quick one. Before you jump back in, you're listening for ways to help teams move faster, make sharper decisions and turn scattered context into work they can use. ChatGPT for Business can help. ChatGPT for Business gives teams a shared workspace with admin controls, permissions and access to work and codecs. In ChatGPT, this means your business can move from question to answer and code to rollout quicker. Join over 10 million business and enterprise users worldwide already using ChatGPT for work. Download the ChatGPT desktop app or contact Sales to learn more.

[Gabriel Hunter]

You probably think of Walmart as the

[Jacob Goldstein]

store that carries just about everything under

[Gabriel Hunter]

the sun and maybe not as the place to discover small brands. Well, two things can be true at the same time, because Walmart is home to thousands of small brands founded by

[Jacob Goldstein]

people who had an idea, took a

[Gabriel Hunter]

chance and built something of their own. Behind every one is a real story and a lot of hard work. So why not take a little time to recognize all the people building small brands across the country and support everything they're creating? Walmart is proud to give those brands a place on its shelves and online. It's never been easier to find and support small brands. So take a look at Walmart. You might come across something new, something unexpected, or a brand whose story speaks to you. You never know what you're missing until you look. Discover thousands of small [e-mail] this summer

[T-Mobile Advertiser]

there are 250 reasons to join T

[Podcast Host]

Mobile Girls Trip booked and switching to T Mobile Help make it happen with

[T-Mobile Advertiser]

experience beyond switch and get $100 back plus $650 in value in built in plan benefits. That's $750 in your first year guaranteed.

[Podcast Host]

And honestly, that hundred dollars back, I'm putting it towards the zipline now.

[T-Mobile Advertiser]

That's how you do summer.

[Podcast Host]

Exactly. Best switch I ever made.

[T-Mobile Advertiser]

This is just one reason to join T Mobile this summer. There are 250 reasons. Find yours.

[Legal Disclaimer Speaker]

Savings based on the value of benefits included with experience beyond like entertainment and one year of AAA classic and dash pass on US benefits may require activation C plan for details get $100 via virtual prepaid MasterCard with eligible ported, no cash access and expires in six months. Issue by sunrise banks and a member FDIC allow eight weeks.

[Jacob Goldstein]

Okay, now we're going to do the lightning round. The questions are going to be somewhat different. What was the hardest thing about writing a play?

[Dr. Sourabh Jha]

It was accepting that nobody would like it.

[Jacob Goldstein]

Did anybody like it?

[Dr. Sourabh Jha]

Oh, they loved it. I was pleasantly surprised. And so it was about a doctor in a rural setting that was trying to stay away from technology, from information and got confronted with a snakebite patient who ended up dying. And it was in rural India where they still blame doctors when things go bad. They treat doctors like God when things go well, but they treat doctors like shit when things go bad.

[Jacob Goldstein]

Do you think AI will replace writers or radiologists?

[Dr. Sourabh Jha]

First it'll replace mediocre writers before it replaces mediocre radiologists. You know the thing is that. Do you read the New York Times or the Economist?

[Jacob Goldstein]

Yes.

[Dr. Sourabh Jha]

Do you know who the author is? And the Economist?

[Jacob Goldstein]

Only for the Obitz because I love her so much.

[Dr. Sourabh Jha]

Yeah, I have trouble because they all sound the same. They all sound the same.

[Jacob Goldstein]

By design. By design.

[Dr. Sourabh Jha]

By design. Because there's a method of writing that they all encompass and AI definitely should take over that. I remember reading people Like Christopher Hitchens. When you would read something by Christopher Hitchens or Salman Rushdie, George Orwell.

[Jacob Goldstein]

Great writers. You are naming great writers.

[Dr. Sourabh Jha]

You know that the prose came from those people. Yeah. Because it was unique and for better or worse, it had a characteristic. And maybe AI will take over that too. Maybe ChatGPT will be able to take that too. But if it comes to uniformity, it should definitely chew that up for breakfast. And I hope it does.

[Jacob Goldstein]

What's the most beautiful scan you've ever seen?

[Dr. Sourabh Jha]

I'd say that the most beautiful scan, unfortunately is of a patient that had multiple pathologies. So it was beautiful on an aesthetic sense, but not so much for the patient sense. It was somebody that had a stroke from a pulmonary embolus, from a clot, what's called a paradoxical embolus, crossing from the right side of the heart to the left side. And as it was crossing over, it was caught in the act. Crossing over, you saw not just the

[Jacob Goldstein]

effect but the cause.

[Dr. Sourabh Jha]

Exactly.

[Jacob Goldstein]

Happening.

[Dr. Sourabh Jha]

Exactly, yeah.

[Jacob Goldstein]

I've seen you mention a book you have called Atlas of Normal Roentgen Variants

[Gabriel Hunter]

that may Simulate Disease.

[Jacob Goldstein]

Tell me about that book.

[Dr. Sourabh Jha]

It's a book that I started reading when I was a resident 20 years ago and I still haven't finished Keats. Not the poet, Keats, the radiologist. Keats compiled a whole range of normal on X rays and novel is really the most difficult thing to. To learn.

[Jacob Goldstein]

Say a little bit more like why

[Gabriel Hunter]

is normal so hard?

[Jacob Goldstein]

It is interesting and not immediately obvious why that is so.

[Dr. Sourabh Jha]

Because there's so many variations of normal. It's almost like what Tolstoy said about happy families all being the same and every unhappy family being unhappy in their own way. You reverse that disease is clear cut. Disease kind of is very similar, homogeneous to some extent, whereas normal has a lot more variability.

[Jacob Goldstein]

What's your view of healthy people getting a full body scan?

[Dr. Sourabh Jha]

Don't bother. Yeah.

[Jacob Goldstein]

Why?

[Dr. Sourabh Jha]

Because you will end up calling over. Calling normal over, diagnosing normal. I mean, yeah, there'll always be the success story of one person whose life was saved and genuinely so. But for every success story there are about 99 where they've been just taken down, you know, a rabbit hole for no reason whatsoever and reassured at the end of it and felt like there's some sort of cancer survival. No, I wouldn't bother. But I also think that it's, it's, it's a decision made by the individual. Some people are risk averse and you have to just take what you risk. The risk tolerances as well.

[Jacob Goldstein]

Anything else we should talk about?

[Dr. Sourabh Jha]

Well, I, I think you've covered a lot. I think that people view AI with both. On one hand you've got this view largely coming from Silicon Valley, and I understand why they need to make money with the view of AI as just being this sort of transforming technology. And on the other hand, there's a fear and skepticism of it. And I think the truth is somewhere in between that it is a technology that needs nurturing and it's a very personal technology. I think it'll end up being very personal to the person nurturing it. No two people will view AI in the same way.

[Jacob Goldstein]

I appreciate your time. Thank you for being so generous with your time.

[Dr. Sourabh Jha]

Thank you very much.

[Gabriel Hunter]

Sourabh Jha is a radiologist at the University of Pennsylvania.

[Jacob Goldstein]

Please email us at problemushkin fm. Tell us what kinds of shows we should do more of or less of or particular people you think would be good on the show.

[Gabriel Hunter]

You can also find me on x and on LinkedIn.

[Jacob Goldstein]

Our show is produced by Gabriel Hunter

[Gabriel Hunter]

Chang and Trina Menino. Our editor is Lydia Jean Cott and our engineer is Sarah Brugiere.

[Jacob Goldstein]

I'm Jacob Goldstein and we'll be back

[Gabriel Hunter]

next week with another episode of what's yous Problem?

[Sponsor Announcer]

Before you sign off, you tuned in for ways to help teams move faster, make sharper decisions, and turn scattered context into work. They can use ChatGPT for business can help. ChatGPT for business gives teams a shared workspace with admin controls, permissions and access to work and codecs in ChatGPT. This means your business can move from question to answer and code to rollout quicker. Join over 10 million business and enterprise users worldwide already using ChatGPT for work. Download the ChatGPT desktop or contact sales to learn more.

[Gabriel Hunter]

When you think about discovering small brands, what store pops into your mind?

[Jacob Goldstein]

Well, it should be Walmart.

[Gabriel Hunter]

Seriously. Walmart has thousands of small brands and they're all in one place. Just go online or in store, discover and shop. It could not be easier. Every one of these brands has a real story and real people behind it. They're true American success stories and you can find them all at Walmart. Discover thousands of small [e-mail] today.

[Jacob Goldstein]

Did you ever notice how you spend

[Gabriel Hunter]

hours shopping online only to pause at

[Jacob Goldstein]

checkout because you wonder if you trust

[Gabriel Hunter]

it enough to hit buy?

[Jacob Goldstein]

Now Agentic Commerce is testing that moment more than ever.

[Gabriel Hunter]

That's where PayPal comes in.

[Jacob Goldstein]

With 25 years of checkouts, 400 million

[Gabriel Hunter]

consumer accounts globally and the benefit of fraud protection.

[Jacob Goldstein]

So no matter where a purchase starts,

[Gabriel Hunter]

it ends with trust. Built for payments, growth and agentic PayPal open built for all business visit paypalopen.

[Legal Disclaimer Speaker]

Com.

[Podcast Host]

This is an iHeart podcast.

[Sponsor Announcer]

Guaranteed human.