Table of Contents
Peter Lee never planned to become a healthcare revolutionary. The Microsoft executive stumbled into medical AI almost by accident, but what he discovered changed everything about how we might think about the future of medicine.
Key Takeaways
- Personal medical documents become instantly understandable when you ask GPT-4 to explain blood test results and insurance paperwork in plain language
- GPT-4 already performs at superhuman levels for "curbside consultations" - those quick questions doctors ask each other throughout the day
- The technology works more like having a brilliant intern than a perfect computer, complete with reasoning abilities but also occasional mistakes
- Healthcare organizations like Mercy and Epic are quietly integrating AI to reduce doctors' clerical burden by multiple hours daily
- Patients actually prefer AI-generated follow-up notes because they feel more human and personal than doctor-written ones
- The medical community needs to take control of AI adoption standards before regulators impose restrictions that could hamper innovation
- We're approaching a future where patients will demand disclosure if their doctor doesn't use AI assistance, reversing today's concerns about AI usage
- Real-world medical experiences could soon contribute directly to advancing medical knowledge through AI synthesis of clinical encounters
The Accidental Healthcare Revolutionary
Peter Lee's path into medical AI reads like a series of fortunate accidents. After writing a two-page policy paper for the Obama administration, he found himself abandoning his Carnegie Mellon professorship to work at DARPA. That first week driving from Pittsburgh to meet the Secretary of Defense, he got pulled over for speeding and downloaded a crowd-sourcing app called Trapster. That simple moment gave him the insight to discuss how network effects and machine learning could transform defense applications.
Years later, when Microsoft CEO Satya Nadella asked him to focus on healthcare in 2016, Lee's initial reaction was pure confusion. "I thought I was being punished for something," he recalls. "What do I know about healthcare? Furthermore, there are 12 other corporate VPs at Microsoft doing their own thing in healthcare - they're not going to listen to Peter Lee."
But sometimes being in the right place with curious people leads somewhere unexpected. Lee discovered that AI, particularly GPT-4, wasn't just another software tool - it was something entirely different that could reshape how we approach medical care.
When Your Blood Work Finally Makes Sense
Here's something most of us can relate to: you get a physical exam, and a week later an email arrives with a PDF full of incomprehensible lab results. Even Lee, an elected member of the National Academy of Medicine, admits he couldn't decode his own blood tests. "You don't feel good about calling your doctor to waste his or her time with this," he explains.
That's where GPT-4 changes everything. Lee simply asks the AI through Bing or ChatGPT to look at his results and explain them. "Is there anything I should be concerned about? My LDL looks a little out of range - what does that mean?" The AI can parse medical jargon and translate it into understandable language, making healthcare information accessible for the first time.
The same applies to those baffling "explanation of benefits" notices from insurance companies. Lee learned that even insurance company executives struggle to understand these documents. GPT-4 can examine those cryptic CPT codes and explain exactly what happened: "Someone in your family had this lab test done, this much was covered by insurance, this much is being covered by your provider, and you don't owe money."
But the technology proved most valuable during a personal crisis. When Lee's father battled illness for 18 months, the family faced a familiar struggle - coordinating care from hundreds of miles away, managing multiple doctors and specialists, and trying to make the most of precious 15-minute phone consultations. Family tensions would flare over how best to use those brief conversations with specialists.
Lee discovered he could feed all the lab results and medical notes to GPT-4 and ask: "We have a 15-minute phone call with Dr. K - what are the best three things to ask and talk about?" The AI's suggestions brought the emotional temperature down, made the family feel more prepared, and actually helped the doctor by ensuring focused, productive conversations.
The Doctor's New Digital Colleague
When Lee demonstrates GPT-4's medical capabilities, the results are striking. He poses a realistic scenario to the AI: a patient with chest pain, elevated heart rate, normal EKG, but elevated troponin levels. GPT-4 immediately recognizes the potential cardiac event and suggests additional tests like coronary angiography and echocardiogram, while noting that other factors like kidney issues could also cause elevated troponin.
When Lee mentions hearing lung rales during examination, GPT-4 escalates its urgency assessment, recommending immediate angiogram and suggesting additional tests like BNP for heart failure assessment and chest X-ray for lung evaluation. The AI demonstrates what Lee calls "almost superhuman" ability to serve as that crucial second set of eyes, critiquing medical reasoning and suggesting overlooked possibilities.
These "curbside consultations" - informal questions doctors ask each other throughout the day - represent one of GPT-4's strongest medical applications. Many doctors work in isolation without easy access to specialist colleagues, making AI consultation invaluable for differential diagnosis development and case review.
Charlotte's Web and the Nature of Machine Intelligence
To explain GPT-4 to healthcare leaders, Lee uses an unexpected analogy from children's literature. He asks audiences if they've read Charlotte's Web (virtually everyone raises their hand), then poses three increasingly complex questions about the book.
First, he asks for a simple character description - easy enough for humans and AI alike. Then comes the harder question: "What does the book try to teach us about the value and nature of friendship?" This requires reading between the lines, making connections to personal experience and social context. GPT-4 handles this thoughtfully, discussing how the book shows friendship as transformative, involving sacrifice and loyalty.
But Lee's masterstroke comes with an original connection. He mentions that the interviewer once led DoSomething.org and asks GPT-4 to find connections between Charlotte's Web's friendship lessons and that organization's mission. The AI draws parallels between empowering young people to create community change and how Charlotte and Wilbur's friendship brings positive transformation - a connection that likely never existed in any training essay.
"I don't think there are many essays on the web about the connection between DoSomething.org and Charlotte's Web," Lee notes. This demonstrates genuine reasoning rather than sophisticated pattern matching.
However, when asked to recite Chapter One word-for-word, neither humans nor GPT-4 can perform. This reveals the technology's nature - it's not a perfect memorization machine but something closer to a reasoning engine with both capabilities and limitations.
The Regulation Dilemma and Medical Leadership
The regulatory landscape for medical AI presents unique challenges. Traditional software medical device frameworks don't apply to technologies like GPT-4, and regulators worldwide hesitate to impose restrictions that might hamper national competitiveness in this crucial technology sector.
Lee argues that the medical community itself must take control of defining when, where, and how AI should be adopted in healthcare. "For that to happen, it's so important for us as a technology community to do everything we can to educate and get the medical world up to speed," he emphasizes.
The disclosure question illustrates how quickly perspectives will shift. Today, we debate whether doctors should disclose AI assistance to patients. Lee predicts that within just a few years - "the distant future in AI is like 2025" - patients will demand disclosure if doctors don't use AI assistance, questioning why their physician didn't double-check diagnoses with artificial intelligence.
He compares this evolution to hand-washing in medicine - something that eventually became an obvious standard practice. "We will definitively get to a point, probably sooner than we think, where of course you would never ever dream of practicing medicine without the assistance of AI."
Lee suggests interim guidelines with set timelines while the technology and its applications mature. The analogy he offers is copper wire - we've discovered it can carry electricity efficiently, knowing it will change everything for the better, but we haven't yet invented the light bulb to demonstrate why. Meanwhile, people are getting electrocuted because we haven't developed proper safety standards.
Real-World Innovation in Healthcare Systems
Microsoft's partnerships with organizations like Mercy Health System and Epic illustrate how AI integration is happening now, not in some distant future. Mercy, a large Southeastern health system with significant in-house engineering capabilities, faces the same nursing shortage crisis affecting the entire industry - over 50,000 nurses short in the US today, potentially millions short within five years.
Epic, the largest electronic health record provider, has integrated GPT-4 into their "Inbox" system that helps doctors communicate with patients. Previously, physicians complained about spending overwhelming amounts of time rummaging through 5-15 or more documents across Epic's system to synthesize useful notes for patients.
Now GPT-4 reads all relevant information and proposes draft communications that doctors can review and edit. The results surprised everyone: patients explicitly prefer these AI-generated notes, saying they feel more human than doctor-written communications. The AI has the "tireless ability" to include personal touches like congratulating patients on becoming grandparents or wishing them well on upcoming weddings - details that busy doctors often miss.
These implementations are reducing doctors' clerical burden by multiple hours daily, representing what Lee calls "the first light bulbs coming on from the copper wire."
The Hallucination Challenge and Human-AI Partnership
The elephant in the room with AI medical applications is hallucination - when the system generates plausible-sounding but incorrect information. Lee embraces the analogy of GPT-4 as your "personal intern" - brilliant and capable, but requiring supervision and verification.
"If you're going to have your personal intern and you're going to have a conversation about your medical issues, you're going to have to make your own assessments about the correctness and veracity of these things," Lee explains. You'll need to probe, ask follow-up questions, and potentially seek additional consultations.
However, he sees significant improvements coming through "grounding" - giving AI systems access to tools like search engines, calculators, and real-time information. For most of GPT-4's public life, it has been "put in a glass box, not allowed to touch anything." Only recently has it gained permission to use Bing search, analyze images, or execute code.
Lee expects these grounding techniques will "dramatically reduce the bad hallucinations while preserving the good hallucinations" - meaning informed guesses and imaginative problem-solving capabilities. It's like giving your intern access to a library, encyclopedias, the internet, and a calculator.
The Vision for Tomorrow's Healthcare
Lee's long-term vision extends beyond individual AI assistants to systemic transformation of medical knowledge generation. He points to COVID-19 as an example of how slowly medical discoveries currently spread. In 2020, some doctors discovered that keeping COVID patients in respiratory distress prone (on their stomachs) could help avoid intubation. A few shared this on social media, but it wasn't until late 2021 that official multi-institutional trials verified the approach - almost too late to help, as the virus had already evolved.
"In the future with AI, we ought to be able to do this routinely - every single clinical encounter and every single medical experience should be synthesizable into clinically sound and validated pieces of medical knowledge and practice," Lee envisions. "I think we actually are very close to having the tools we need to make that a reality."
This represents a fundamental shift toward "real-world evidence" where every healthcare interaction contributes directly to advancing medical understanding and treatment protocols.
Looking ahead 15 years, Lee sees unprecedented optimism in technology's potential. He believes the combination of AI's ability to model complex physical systems and the emergence of practical fusion energy could help avert climate catastrophe while creating foundations for technologies that make prosperity accessible to much larger portions of the global population.
The first step toward this future isn't waiting for perfect technology or comprehensive regulations. It's about medical professionals, technologists, and patients beginning to experiment thoughtfully with these tools today, learning their capabilities and limitations through real-world experience. As Lee learned through his own career of accidental discoveries, sometimes the most transformational changes happen when smart people ask you to try something new - and you're curious enough to say yes.
