AI in Healthcare: Dr. Robert Wachter on What’s Next for Providers and Patients

Robert M. Wachter, MD is Professor and Chair of the Department of Medicine at the University of California, San Francisco (UCSF). The author of 300 articles and six books, he coined the term “hospitalist” in 1996 and is often considered the “father” of the hospitalist field – the fastest-growing medical specialty in US history. He is a past president of the Society of Hospital Medicine, past chair of the American Board of Internal Medicine, and an elected member of the National Academy of Medicine. In 2004, he received the John M. Eisenberg Award, which is the nation’s top honor in patient safety. Modern Healthcare magazine ranked him among the 50 most influential physician-executives in the US more than a dozen times, and he topped the list in 2015. His 2015 book, The Digital Doctor: Hope, Hype and Harm at the Dawn of Medicine’s Computer Age, was a New York Times bestseller. His new book, A Giant Leap: How AI is Transforming Healthcare and What That Means for Our Future, was an instant national bestseller and is being developed into a documentary.

Dr. Wachter will present a keynote address at the 2026 IHI Forum (December 6–9, 2026 in Phoenix, AZ).

Image

What are the biggest missed opportunities right now for using AI in healthcare?

The healthcare system is massively broken, and therefore, there are improvement opportunities everywhere. 

There are opportunities for AI on the administrative side, given the amount of paperwork that doctors, nurses, and other providers need to handle, as well as the amount that patients need to wade through. We’ve started with AI scribes and appointment scheduling, but healthcare is still in the last century. 

And there are opportunities on the knowledge side as well. As a general internist, as a result of AI, I now have access to subspecialty-level knowledge in every field. That doesn’t mean I never need to call a cardiologist or infectious disease specialist, but it does mean that if I have a question where I would normally hope to run into someone in the hall to ask, now I can have it answered by AI.

Patients and caregivers have access to more information than ever before, thanks to AI. There may be times when they don’t need to access the healthcare system to have their questions answered. 

AI should improve the experience for everyone, in theory, but whether it will in practice is a harder question. As a result of my research, I’m hopeful that the benefits in healthcare will far outweigh the downsides.

In what situations do you least want to see AI used in healthcare?

The things that make generative AI so magical and remarkable are a double-edged sword. With the opportunities come threats.

The biggest downsides are less around individual uses and more the fact that AI is a spectacular potential promoter of misinformation. And, as we depend more and more on AI — whether it’s offering decision support to a clinician or embedded in your insulin pump — if these systems get hacked, there’s more mischief that can occur. 

In terms of clinical decision support, when I use AI as a trained professional, I find it incredibly helpful. It’s almost always right. It hallucinates far less than it did three or four years ago.

However, the early studies of patient use are a little bit sobering. The patient may not know what the most important information is to put into the tool. For example, when I see a patient who’s short of breath, one of my first questions is, are your legs swollen? Someone who understands medicine knows that if one leg is swollen, I need to be worried about a blood clot. If both legs are swollen, I’m worried about heart failure. There’s a risk that patients may not recognize that a certain issue is important or may enter incorrect information. There’s an old saying in informatics, “Garbage in, garbage out.” The AI is only as good as the information that it has.

For a patient whose next available appointment to see a primary care doctor is a month away, or who is in a rural community without access to the relevant specialist, I think the answers from AI are much better than nothing. Patient-facing tools are getting better. If you say, “I woke up today with a headache,” you want the AI to do what a good doctor would do, which is to ask questions before rendering an answer: When did it start? Is your neck stiff? Does bright light hurt your eyes? But it’s important that if a patient has severe chest pain, or wakes up to find that one side of their body is weak: Put down the phone and go to the hospital.

The more information the AI has about you, the better it will be. Patients need to make a choice about how much information to give, particularly when they’re dealing with AI companies. I opened an AI tool about a month or two ago and it said, “Do you want to load in your electronic medical record?” It said the company promised not to sell the data, leak the data, or use it for training. But, they don’t operate under the same privacy rules as my hospital or doctor or electronic health record vendor, so I need to decide if the benefits of them having my data outweigh a small privacy risk. 

I chose to move my data over. So now, the AI knows what medicines I’m on; it knows that I fell down a few years ago and had a small bleed around my brain. The more accurate information it knows, the more likely it is to give a useful answer. But, at least for now, it’s a little bit of “buyer beware.” 

What do you see in terms of embedding equity into AI?

While there are risks, it should be an improvement. As to whether it really will be — the proof will be in the pudding. The tools can know a lot about the user: their preferred language, their health literacy, what their insurance will and won’t pay for. Armed with that knowledge, the tools can provide behavioral counseling that is somewhat customized.

The greatest boon for equity would be for patients who don’t have access to professionals. For example, in mental health, we’ve seen real problems with some of the early tools — cases in which they seemed to counsel teenagers to harm themselves. On the other hand, every day millions and millions of people are getting mental health help that they find useful, free or for $20 a month, whereas finding a mental health professional in their town may be impossible or might cost a few hundred dollars per hour.

The capacity of tools to scale biases is also real. All the algorithms know is what they learn from medical literature. There are studies that show that if two patients came to an emergency room with exactly the same fracture, and one was White and one was Black, the White patient would get more pain medicine than the Black patient. AI might look at that and conclude that White patients need more pain medicine. So scaling that bias is risky. However, there’s also an opportunity to measure that bias and even correct it. In some ways, it’s easier to change what machines do than what humans do.

Another issue around equity is whether people will have what they need to use AI. For example, can they afford the subscription to an AI tool that works? Do they have wireless internet access?

Overall, I would say this will improve equity, in part by democratizing expertise, but it’s an area we need to watch carefully.

Interesting. That sounds hopeful.

After two years of research and 110 interviews for my book, I landed in a pretty optimistic place, at least in healthcare. But it’s crucial to realize that we have a Category Five AI backlash going on. I have many of the same worries as others — about hacking, climate, jobs, education, and income distribution. On the other hand, healthcare is in such desperate need of improvement that the status quo is immoral. It’s unacceptable. AI represents our best chance of improving the status quo that I’ve seen in the last generation.

What advice do you have for healthcare professionals who are struggling with uncertainty about the future of AI?

I don’t see how you can be the best health professional you can be without using these tools. AI scribes can allow you to look the patient in the eye rather than being a data entry clerk; AI tools can accurately summarize a patient’s 800-page chart; AI can suggest a diagnosis that you might not have thought of; AI can allow your patients to do things for themselves that previously would have required a visit to the doctor. To me, without question, these tools make me a better doctor. And, when I use it myself or family members use it, I think it makes us more informed and capable patients. 

But there are very real concerns, even in healthcare. AI will lead to fewer administrative jobs. That’s sad for people who might lose their jobs, but administrative bloat has been a major contributor to runaway healthcare costs. On the other hand, I don’t think doctors or nurses are going to lose their jobs. Even if these tools allow patients to manage their blood pressure and cholesterol without seeing a clinician, there’s still so much unmet need. 

When it comes to AI in healthcare, we often talk about the “human in the loop” — a health care professional who is reviewing and approving the AI’s work. What’s that experience like for the human?

The “human in loop” model says, if the AI is good enough to be useful, but not perfect enough to be fully trusted, then have a human look at the AI’s recommendation and then endorse it, edit it, or reject it. That sounds like a safer system than it is. Part of the problem is de-skilling — over time, humans will become less proficient at the task if they become overly dependent on the AI. 

More importantly, the AI-human dyad is often not a robust patient safety system. None of us is good at remaining eternally vigilant when we’ve come to trust a technology. If the tool has been right 20 times in a row, few of us will be really careful and thoughtful the twenty-first time. 

There’s another risk: If the AI gets better, the human in the loop may actually make things worse. If the AI is right 95 percent of the time and the human is right 75 percent of the time, on average, the human will reduce the dyad’s performance. We’re going to have to study how the human-AI pair works together and what kind of performance they achieve.

If we do want the human in the loop, what kind of system can we build to make it more reliable? For example, at the airport, the TSA system periodically adds an image of a bomb or a gun into the video feed to be sure the agents stay on their toes. You can imagine in healthcare having the AI periodically give a wrong answer on purpose. Companies are also working on tools that share the level of confidence in a particular answer, to signal to the human when to pay more attention.

The state of Utah is studying an AI tool that’s been granted permission to refill certain types of medications without a human in the loop. With no human in the loop, the company needed to buy malpractice insurance for their AI. 

We need to ask ourselves, are there tasks that can be done effectively, safely, more conveniently, and less expensively by AI? There are still enough human tasks that require compassion and empathy. I’m fine with AI refilling my cholesterol medicine. I'm not fine with AI telling me that my kidneys are failing or I have diabetes or cancer.

Recently, you wrote that we need to move quickly when it comes to adopting existing AI technology in healthcare, as opposed to pushing the AI frontier quickly. Can you describe that difference?

Much of the backlash against data centers, AI hacks, and so on, relates to a concern that we’re going too fast. If we lose control of these AIs, and they begin to hack into the electrical grid, we have a big problem. I think that’s a different question than whether we should move fast with existing AI tools in healthcare. 

Even if there were absolutely no new models developed by AI companies in the next three years, it’s going to take us that long to absorb what’s already out there. It’s not a matter of implementing an AI tool and switching it on. You have to think about its impact on the workflow, the people, the culture, the payment system, and so on. There's a lot to do to get this right. 

It seems reasonable to me to go a little slower with frontier AI models, given concerns about the tools escaping human control and causing harm. When it comes to implementation in healthcare, we have a lot of natural regulators that already keep things slow, such as the regulatory system, malpractice suits, and professional culture. In healthcare, I think the risks of going too slow are higher than the risks of going too fast.

You might also be interested in:

• Leading with Love: Katie Boston-Leary on Advocacy, AI, and Advice for Addressing Burnout
• What Art Can Teach Healthcare About True Collaboration
• The Potential and Pitfalls of Using AI for Safety Event Review