This week three nationally recognized physician-AI experts weighed in on where they think AI deployment in healthcare is likely to go in the next five to 10 years.

The panelist took part in a seminar co-sponsored by The Physicians Foundation Center for the Study of Physician Practice and Leadership and the Cornell Health Policy Center, a collaborative partnership between Weill Cornell Medicine and Cornell University’s Brooks School of Public Policy. The conversation was moderated by Dhruv Khullar, M.D., M.P.P., a physician and associate professor of health policy and economics at Weill Cornell Medical College.

Khullar first asked the panelists what they think people should be paying more attention to with AI and conversely what they're paying too much attention to right now.

Adam Rodman, M.D., M.P.H., director of AI programs at the Shapiro Center for Research and Education at the Beth Israel Deaconess Medical Center, said he thinks people are paying way too much attention to some of the clinical reasoning skills of generative AI in a way that attracts headlines, but may not be very clinically useful. “I think that the findings of some of my own research are probably overhyped,” Rodman said. “The abilities of a model to perform some task are not the same as making a workflow that makes a difference in a patient's life, and there are many real socio-technical barriers to putting these technologies out.” 

Rodman added that we're in a startup ecosystem where, because people will throw money at any company that says AI doctor, there's a lot of hype, and it takes up a lot of airspace. “The reality is that the technology being able to do something cognitive is not the same as making a difference in the real world.”

Khullar asked what will medicine look like in 5 to 10 years that's tangibly different than it is today.

Roxana Daneshjou, M.D., Ph.D., assistant professor of biomedical data science and of dermatology at Stanford University School of Medicine, said that unfortunately she has kind of a dark vision of where this is going and it has nothing to do with AI but more to do with forces at large.

She noted how difficult it is in many parts of the country to get any kind of primary care, and it’s even harder to see specialists. “At the same time, we're seeing cuts in reimbursement, which is also contracting what kind of patients with what kind of insurance physicians might see,” she said. “Those headwinds are part of the reason that we see a lot more direct-to-consumer companies like Function Health. These direct-to-consumer companies are rising in power because people are so desperate to get some kind of healthcare. They're turning to AI models to look at their healthcare because it's taking forever to see a primary care doctor. It might take six to nine months to see a specialist, and they're having very real debilitating symptoms and can't get in.”

In addition, Daneshjou said, consolidation is happening in parts of medicine where private equity is buying up practices and trying to reduce costs. She expects that we’ll see more mid-level clinicians with AI replacing physicians. It is not because necessarily mid-levels plus AI is better than a physician or even AI itself, it's because the healthcare system is buckling, she added. “The people who are actually making the decisions and are in the C-suite are not physicians; they're businesspeople, and they're going to make decisions that are right for their business, even if it's not necessarily the best thing for the patient. I think all these different headwinds coming in are leading us to this conversions where AI tools plus someone who is maybe not as well trained will be the way that many people get care in the next 10 years.”

Khullar said he appreciated that many forces are shaping medicine and will shape how AI is used in medicine, but he asked: Is it a bad future where clinicians are extended in this way? “And if so, is it because the technology is or is not going to be good enough to develop into the system that delivers high-quality care, or is there something else that we should be keeping in mind? Some people could argue that that's a good system. We have a primary care shortage; now we have all these clinicians who are armed with various types of AI, and that could go a long way.”

Daneshjou, who also is assistant director of the Center of Excellence for Precision Heath & Pharmacogenomics, responded that the current state of things, where people can't get preventative care and the care that they need, is not a good state. So if AI caused more people to be able to get the quality care that they needed, and it was better than the current status quo, that’s a good future. But that may not be the optimal future. “If we wanted to be honest, to go toward the optimal future, we would have more primary care doctors. We would actually put more money into public health. I am an AI scientist, but in my mind, there are investments you could make in vaccinations and helping with childhood diseases and nutrition in childhood that would probably save far more lives than all the money that we're putting into AI.”

Jonathan H. Chen, M.D., Ph.D., associate professor of medicine and of biomedical data science at Stanford and director for medical education in artificial intelligence in the Stanford Division of Computational Medicine, said he thinks we overestimate and overhype what's going to happen in the short term and we probably are underestimating what's going to happen in the longer term. He said 95% of the startups in the space will be dead in five years. “I'm very confident about that, but a few of them will probably emerge as the most dominant companies in the world, and also dramatically change what we even expect medicine to be,” he said. 

Asked what will be different, Chen said that autonomous prescribing is going to happen. “I would not have even said that out loud three years ago because it would have sounded scary. It would have made people angry. Now it's obvious. It’s kind of already started to happen,” he said. 

Chen noted that some states are trying to ban the use of AI for mental health delivery. “I get it, because there are some really scary stories in the news about where this can go way off the rails. But half the people who can benefit from therapy don't have access. They can't get to somebody, let alone afford the co-pay. So it's not like our status quo system is great either. And it's not because the robots are so great. It's that humans are going to naturally want to reach for what they can at least get their hands on.” He said there is a lot of good AI can do while health systems do their best to mitigate harm.

Rodman added that he thinks that in 10 years we'll start to really see the impacts of patient-facing conversational AI, noting that the payment system matters so much. “If we stick with our current payment system, we're going to get a pretty dark future where efficiency gains just make my 20-minute appointments become 10-minute appointments. But regardless, I think patient-facing AI is going to be a huge part of healthcare 10 years from now. But he added that some of these patient-facing technologies could actually be inflationary. “I think a lot of the assumptions that it saves money are probably not true. Now, whether it makes people healthier is another question, but I don’t think patient-facing AI is going to save money.”

A two-tiered system?

Khullar asked if there is a potential danger of this leading to a two-tiered health system where people who can afford human care will get it and other people will have some combination type of care. He said another dimension on which this could be two-tiered is that there are various forms of transactional care. “You have a UTI, you have a sprained ankle, pretty straightforward things that you can converse with a chatbot and get a prescription for or imaging study for. But for complicated, complex care, let's say for advanced heart failure or COPD or certain types of cancers, that will still be primarily human-first. Do you buy that?”

Daneshjou responded that we already have a two-tiered system. “The investments that we're making around AI currently are not pushing us anywhere away from an already existing two-tiered system,” she said. “A lot of my research has been in AI bias, so I am deeply concerned about basically encoding the existing biases of the healthcare system, whether that's by race, by gender, or any other sort of protected attribute into the systems. And in fact, in our testing of models, we have seen that that's already happening. That's a huge concern and something that I urge people who are developing these and testing these AI systems to check for. It's something that our team always checks for.”

She said she cannot say whether complex care that is human-led or AI-led is going to be better. Daneshjou noted that there is the average doctor, and half will be better than average, and half will be worse. Even at the top tier of complex care, that is true. “So then the question becomes: On average, will AI in a decade do better at complex care and be more standardized? Maybe. I would need to see the data to show that, but I think it's possible. And looking at the last decade, it's quite possible that that could happen in the next.”