Imagine that the only way Americans traveled was on foot or on horseback. And assume that 100,000 people died each year because they couldn’t reach a hospital in time or firefighters arrived too late.
Suddenly, they learned that thanks to a technological breakthrough, cars and trucks will become widely available within three years.
How would our nation’s leaders respond? I predict that elected officials and regulatory agencies would move quickly to build roads, accelerate automotive production, and help people learn to drive, rather than waiting for cars to become 100% risk-free. They’d recognize the nation’s death toll would drop significantly with faster transportation.
This is how American leaders should think about generative AI in medicine.
Medicine’s deadly status quo
Most Americans know healthcare is too expensive and difficult to access. Far fewer understand that hundreds of thousands of people die annually from misdiagnoses, poorly controlled chronic disease, and preventable medical errors. Fewer still can imagine how large language models like ChatGPT, Claude, and Gemini would help prevent those deaths.
Today, diagnostic errors kill or permanently disable nearly 800,000 Americans each year. In addition, chronic diseases like hypertension and diabetes remain poorly controlled for millions of Americans, contributing to at least 50% of the nation’s heart attacks, strokes, and kidney failures each year. And nearly 60% of patients with a family doctor say they can’t get an appointment for weeks.
GenAI’s lifesaving trajectory
When the first generative AI tools reached the public in November 2022, the technology was deemed remarkable but unreliable, and far from safe in clinical settings. Google’s first medical AI model, Med-PaLM, achieved a mere 67% on the U.S. medical licensing exam later that year. Researchers, clinicians, and regulators quickly urged caution. But by March 2023, Med-PaLM 2 had scored 87%: expert doctor level.
Today’s models are even more impressive. A new Harvard-led study tested OpenAI’s o1 preview model on difficult clinical tasks, including 76 real emergency-room cases at Beth Israel Deaconess Medical Center in Boston. At three stages of care (initial triage, first physician contact, and hospital admission), the model matched or exceeded the performance of experienced physicians.
Assuming large language models continue to double in power over the next three years, they will become 8 to 16 times more clinically capable than they are today. And yet, healthcare practitioners and leaders continue to view GenAI’s capabilities as limited to administrative tasks such as billing, coding, documentation, and ambient listening. Used to improve clinical outcomes, AI could help prevent hundreds of thousands of deaths each year.
Releasing the brakes
GenAI’s next set of lifesaving tools is already within reach. Existing models are being trained on ever-richer clinical information: real conversations between doctors and patients, millions of medical records, and streams of data from hospital monitors (97% of which goes unanalyzed because there is too much information for humans to process).
Given the toll of lives lost from today’s medical failures, the safest choice for our nation is to move our foot from the brake to the accelerator. Here are three examples where GenAI can help address today’s failures:
1. Chronic disease control
Conditions like hypertension, diabetes, and heart failure affect 75% of American adults and produce many of the nation’s deadliest complications, including heart attacks, strokes, and kidney failure.
Despite effective medications and evidence-based protocols, these diseases remain inadequately controlled. Two major reasons are medical structure and clinician time.
Across the country, chronic diseases are managed episodically, with office visits lasting 18 minutes on average, scheduled three or four months apart. Instead, patients need frequent monitoring to identify clinical trends and allow clinicians to adjust medication quickly when effective control isn’t happening.
A GenAI application connected to a home blood pressure cuff, glucose monitor, or wearable device could continuously track whether a patient’s condition was improving or worsening. Patients would continue their current plan, while doctors would be notified of problematic trends and able to make timely medication adjustments.
2. 24/7 medical guidance
When people develop symptoms at night or on weekends, they must decide whether to wait until the doctor’s office opens or go to an overcrowded emergency department.
Generative AI tools could help people select the best option by providing personalized guidance on the most likely diagnosis, the appropriate treatment, and whether to seek immediate medical care. Often, the problem resolves without additional medical assistance.
3. Ongoing clinical support
Most misdiagnoses and preventable medical errors happen not because clinicians lack knowledge, but because they lack the time to apply it.
During brief office visits, they are expected to evaluate new problems, manage chronic conditions, address preventive care, and comfort anxious patients. It’s not possible.
Technology companies are already building AI tools to reduce administrative burden. They are useful, but insufficient. GenAI could also provide immediate expertise to individuals with straightforward medical problems, freeing clinicians to spend more time with those patients at greatest risk.
The government’s role
To translate these opportunities into practice, the nation needs a coordinated public-private effort.
The federal government needs to ensure that all Americans, regardless of location or income, have access to the most reliable and robust large language models.
NIH researchers, medical societies, and clinicians will need to develop patient education materials that teach people how to prompt GenAI models and ask effective follow-up questions. They would also work with technology companies to ensure LLMs consistently provide reliable, evidence-based answers.
Finally, the FDA and CDC need to ensure that wearable and bedside monitors connect safely with GenAI applications on patients’ phones and computers.
If the United States wants to save the greatest number of lives, now is the time to expand access to GenAI tools, build safe and reliable technological applications for patients, and teach all Americans how to use GenAI to maximize.
Robert Pearl, the author of “ChatGPT, MD,” teaches at both the Stanford University School of Medicine and the Stanford Graduate School of Business. He is a former CEO of The Permanente Medical Group.
