I recently had the chance to sit down with Dr. Resa Lewiss on The Visible Voices Podcast to talk through where we actually are with AI in healthcare—not where the hype cycle wants us to believe we are, and not where the doomsayers warn we’re headed. If you’ve been following our work at HEAL Lab, you know we’re interested in the messy middle: the real implementation challenges, the human factors that determine whether these tools help or harm, and the philosophical questions about what it means to practice medicine in an age of increasingly capable machines.
Listen to the full episode here
The Bird Flu Problem: What LLMs Can and Can’t Do
Resa opened with a perfect example of where we are right now. Researchers at University of Maryland used a large language model to analyze over 13,000 ED visits, searching for high-risk bird flu exposures among patients with respiratory symptoms. The AI identified 14 patients with occupational exposures—butchers, farm workers—that might have been overlooked in the chaos of routine care.
It’s a good use case. It’s also revealing about the limitations.
“The model is great to find the things we might overlook, to piece together notes from all different places, which I just can’t have access to. But then there’s also the element of using the same tools to record my conversation, to let me have a normal human conversation and catch all the stuff that I might have missed.”
Here’s the thing: the LLM can only find what someone wrote down in the first place. If I’m seeing four patients in a row and don’t capture that my ankle sprain patient works in poultry processing, no amount of sophisticated NLP will surface that risk factor later. This is why we’re so interested in ambient scribes at HEAL Lab—not just as documentation tools, but as ways to capture the social determinants of health and contextual details that get lost when we’re cognitively overloaded.
The bird flu example illustrates something we talk about a lot: AI as a safety net for human limitations. We know we’re going to miss things. We know that “common things are common” heuristics sometimes cause us to overlook zebras. An LLM that can scan thousands of notes to flag the 14 high-risk exposures we might have dismissed? That’s valuable. But it’s collaboration, not automation.
The Third Wave: Why This Time Might Be Different (Or Not)
Resa asked me to walk through the history of AI hype in medicine, and it’s worth revisiting because we’ve been here before—twice.
The first wave in the 1960s: Expert systems and recommendation engines. People were programming computers with rules: “If white blood cell count is X, then diagnose Y.” It fizzled because we didn’t have the computational power to handle medicine’s complexity.
The second wave in the 1980s-90s: Finance modeling techniques migrated to healthcare. Could we predict patient outcomes the way we predict markets? Computation improved, but this time the limiting factor was data. Those paper charts couldn’t fuel the models.
“Cut to the third one, post High Tech Act, post us going, okay, we need better data. Those paper notes can never be used for any future research ‘cause no one can find them... The silver lining was that it has provided the basis of us writing our logic, having machine interpretable and human interpretable information to start working off of.”
Now, the third wave: We have the compute. We have the data (thanks, EHR). And we have these transformer models that are genuinely good at pattern recognition across massive datasets. So maybe this time is different?
Or maybe we’re just hitting a different wall. This time, the limitation isn’t compute or data—it’s human factors, implementation science, and our inability to design systems that actually collaborate with clinicians rather than trying to replace them.
If you’ve read our piece on the “777 problem”—using overpowered technology for simple tasks—you know where this is going. The question isn’t whether the technology is impressive. It’s whether we’re deploying it in ways that actually improve care.
The Beetle Box Problem: Why We Need to Talk About AI
Toward the end of our conversation, I brought up something I learned in college: Wittgenstein’s beetle box problem. It’s become a touchstone for how I think about communication, medical education, and now, large language models.
“You understand the world as you see it, but it’s walled off from everyone. And the only way to engage and to help each other understand it is by engaging in communication and using your voice.”
Wittgenstein’s thought experiment goes like this: Imagine everyone has a box with a beetle inside. No one can see into anyone else’s box. We all call the thing in our box a “beetle,” but we have no way of knowing if we’re talking about the same thing. The only way we create shared meaning is through language—through the back-and-forth of human communication.
This matters for AI in two ways.
First, it’s a reminder that LLMs are fundamentally different from human understanding. They don’t have a beetle in their box. They have statistical patterns across billions of words, but they lack the embodied experience, the tacit knowledge, the situational context that shapes how doctors actually think. When an LLM tells you something with apparent confidence, it’s not drawing on understanding—it’s interpolating patterns.
Second, and maybe more importantly: the way we learn to be doctors is through conversation. Medical education isn’t just knowledge transfer. It’s apprenticeship, it’s rounding, it’s case discussions where we externalize our clinical reasoning and get pushed back on by attendings and peers. When I worried earlier about medical schools treating physicians as “just knowledge workers,” this is what I mean. If we reduce doctoring to information retrieval—something LLMs excel at—we lose the collaborative sense-making that defines expertise.
What Gets Lost When We Automate
There’s a tension running through the whole conversation that I want to name explicitly, because it connects to everything we’re working on at HEAL Lab.
Automation promises to eliminate the need for human expertise. You don’t need to know how to shift gears anymore because your transmission is automatic. You don’t need to navigate because Google Maps does it for you.
Collaboration tools amplify human expertise. A stethoscope doesn’t replace clinical judgment—it extends it. A chain saw doesn’t eliminate the need for a skilled carpenter.
The question is: Which are we building?
In the podcast, I talk about the Air France 447 disaster—pilots who had become so dependent on autopilot that when the sensors froze and the system handed control back to them, they couldn’t recover. They had lost the expertise they needed to be the fail-safe.
Compare that to heads-up displays in cockpits: pure collaboration tools that keep pilots in the loop, maintaining and building expertise even while providing crucial information. The HUD doesn’t fly the plane. It helps the pilot fly it better.
We’re at a choice point with AI in medicine. We can build systems that try to automate diagnosis and decision-making, accepting the reality that doctors will become deskilled over time. Or we can build systems designed for collaboration—tools that make experts better at what they do, that surface relevant information without overriding human judgment, that keep clinicians in the loop in ways that maintain and enhance expertise.
Go Listen
There’s obviously much more in the full conversation—we get into global health equity, medical education, the risks of chatbots that just tell patients what they want to hear, and why being a middle child from rural New York shaped how I think about all of this.
Listen to the full episode on The Visible Voices Podcast
And if this resonates, you know where to find us: heallabnotes.substack.com for our ongoing work on AI implementation, and stanford-heal.com for the lab.
The beetle boxes are all still closed. But at least we’re starting to talk about what’s inside them.
