Sep 14, 2026 · 20 min · 8 segments
In this episode of SkillCast by LOSD, host Nurbanu Somani speaks with Dr Amish Purohit, a physician, healthcare executive, entrepreneur and Harvard alumnus whose career spans clinical medicine…
Amish PurohitGuest
Nurbanu SomaniHost
But what's really intrigued me from what you're working on is, and what was going through my head was, how is this different to people using ChatGPT or OpenAI or any of these AI sources to help kind of condense and work out treatment plans for patient, con- digest information that the practitioner's having with the patient? Like, what makes it different? Why, why should people not use the automated stuff on OpenAI, for example? Or are you integrating that within your practice?

Yeah, I think that, you know, I, I'll be honest, I think that, uh, ChatGPT, Claude, Perplexity, these AI agents, uh, they're, they're fairly good, right? I think one of the problems that, um, when you work with AI, and I have a postgraduate, uh, program from the University of Texas in AI as well, so I understand the space very well.

The problem with using a large language model, 'cause that's all these are, they're next word predictors.

And so what happens with these large language models is if they're not trained correctly, they will give you information, and they'll give you wrong information sometimes.

And a discerning eye, like a, a, a trained provider, can pick those things out.

Like, if ChatGPT is giving you, um, information that you think is credible but you don't know any better, and I could look at it and it's like, "You're completely wrong.


They'll give you generic information, and sometimes they don't know when to stop, right? And so when they don't know when to stop, it is what they call AI hallucinations, and I would hate for anyone to go and ask ChatGPT, and then they have serious complication because they're like, "Oh, you know, ChatGPT told me this is what I should do." So I would caution, truly caution patients.

It's just like asking Google, "Hey, what do I do for this?" I think it's great to kind of understand, like, if you're gonna put in some symptoms and like, what do you think might be going on, fine, right? But I wouldn't say, "How should I treat this? How should I do this?" Because no one's clinically validated these tools.

Um, for clinicians in the United States, there's a company called OpenEvidence that gives the latest clinical evidence on different topics that you can ask it as a provider.

You can even consult it without using, you know, protected health information.

What do you think might be going on?" I would trust that resource, right? Because it's vetted.

Like, it's okay, but you need a real person, a real human, that input into that AI that's actually giving you much more, a much more clear-

... plan for e- that particular individual, because everyone is individualistic.

A very much overview of someone's symptoms isn't a representation of everything else that's going on with that one individual.

But what's really intrigued me from what you're working on is, and what was going through my head was, how is this different to people using ChatGPT or OpenAI or any of these AI sources to help kind of condense and work out treatment plans for patient, con- digest information that the practitioner's having with the patient? Like, what makes it different? Why, why should people not use the automated stuff on OpenAI, for example? Or are you integrating that within your practice?

Yeah, I think that, you know, I, I'll be honest, I think that, uh, ChatGPT, Claude, Perplexity, these AI agents, uh, they're, they're fairly good, right? I think one of the problems that, um, when you work with AI, and I have a postgraduate, uh, program from the University of Texas in AI as well, so I understand the space very well.

The problem with using a large language model, 'cause that's all these are, they're next word predictors.

And so what happens with these large language models is if they're not trained correctly, they will give you information, and they'll give you wrong information sometimes.

And a discerning eye, like a, a, a trained provider, can pick those things out.

Like, if ChatGPT is giving you, um, information that you think is credible but you don't know any better, and I could look at it and it's like, "You're completely wrong.


They'll give you generic information, and sometimes they don't know when to stop, right? And so when they don't know when to stop, it is what they call AI hallucinations, and I would hate for anyone to go and ask ChatGPT, and then they have serious complication because they're like, "Oh, you know, ChatGPT told me this is what I should do." So I would caution, truly caution patients.

It's just like asking Google, "Hey, what do I do for this?" I think it's great to kind of understand, like, if you're gonna put in some symptoms and like, what do you think might be going on, fine, right? But I wouldn't say, "How should I treat this? How should I do this?" Because no one's clinically validated these tools.

Um, for clinicians in the United States, there's a company called OpenEvidence that gives the latest clinical evidence on different topics that you can ask it as a provider.

You can even consult it without using, you know, protected health information.

What do you think might be going on?" I would trust that resource, right? Because it's vetted.

Like, it's okay, but you need a real person, a real human, that input into that AI that's actually giving you much more, a much more clear-

... plan for e- that particular individual, because everyone is individualistic.

A very much overview of someone's symptoms isn't a representation of everything else that's going on with that one individual.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.