The pace of adoption of generative AI tools in health care has outrun the pace of validation, leaving clinicians to reckon with systems their patients may already be consulting daily. At Cleveland Clinic’s A.I. Summit for Healthcare Professionals, Pete Clardy, MD, Director of Clinical Enterprise at Google for Health, placed recent advances in historical context and discussed their implications for care delivery and medical education.

He explains why AI outputs should be treated as objects for inspection rather than sources of truth, how desperation rather than trust is driving patient use of these tools, and which emerging technologies, including multimodal co-clinician tools, world models, and geospatial data, he believes matter most for health.

Transcript: First of all, I was delighted to have a chance to present this morning at the Cleveland Clinic AI Symposium. A few of the things that I wanted to provide in terms of takeaways were to help put some of the recent developments in artificial intelligence into a more of a historical context, and also to think about the way that these new tools have implications for both how we deliver care and educate future physicians and clinicians.

I think one of the things that's really most interesting about the ways that the work that we do, both inside of Google and the changes that we see across the health ecosystem that have impacted our work, are the degree to which people are using these tools already in their everyday lives to understand their health in greater detail than ever before. And also, the ways in which these tools can be used, not only as a sort of question–answer, call-and-response chatbot experience, but rather how these tools can be used in combination to solve much more complex problems and workflows, both in areas like scientific discovery and also in areas like clinical management.

In health care, we need to establish trust to really have very clear proof points to understand that new technology is not only safe but effective in bringing new value, whether that's saving time or saving lives. That requires trust to build, and that trust requires time and diligence.

I think what I'm balancing that with is that on the person side, or the patient side. I see these tools as leveling historical information asymmetry, information that historically was locked up in medical textbooks or in your medical record is now becoming available for people to learn from and interact with. And I think the dichotomy in terms of trust and desperation comes when you think about how everyone needs an advocate and a wayfinder when it comes to managing their own health.

One way of thinking about these tools is that when people are desperate for wayfinding and for support, they will use new tools in different ways. And I think we in the health-care profession are learning a lot from the ways in which people are using these tools in understanding their health journeys in much greater detail than before. And that's often not driven by trust, but rather by desperation.

The on ramps for a certain kind of creativity and the ability to build things, which historically people could only imagine, has really changed my way of thinking about the work I do or the tools I might build for educational purposes. And so I think that for me, it's been both understanding the limitations of the tools, but also practicing to use new tools in different ways. One of the comments I made in my talk is that it's very easy to get enamored of the technology and the theory, but it's really important to just start practicing safely with whatever tools are close at hand. Because I think until people start to develop a sense of what works and what doesn't, it still remains very abstract. So I think the rapidity with which tools for building have increased access is really surprising to me.

It's really important for clinicians to understand the outputs of AI as objects for inspection, not sources of truth. I think that really helps me evaluate critically things that look very seamless and fluent but are wrong. And I think it's really important that we help people understand that these tools generate outputs that are probabilistic and imperfect and require refraction through human cognition and human context to really become valuable.

So that understanding of the difference between an object of inspection and a source of truth, that's probably the most important learning. A couple of the things that I referenced today that I think are really interesting: One is multimodal, real-time AV co-clinician tools—tools that bring sight and sound together for understanding and communication in a much more natural and human way and start to leverage some of the things that human providers are really good at related to nonverbal communication and also things like physical diagnosis. So I think that area is really interesting and evolving very, very quickly. I also think that some of the things that I mentioned in terms of world models that are not language or multimodal models in the traditional sense that we've understood them, but really models that try and understand the physics of how the world works at large and small scales, is super interesting.

A related but different set of efforts are better understandings of geospatial data. That has an enormous amount to do with health, not only as it relates to weather and natural disasters and things like air quality and pollen counts, but also as it relates to the geospatial distribution of goods and services and the patterns in which people move. All of which provide insight into what's available locally and help to address some of the challenges in health that are separate from the delivery of medical care and much more about the environment in which people live.

*Transcript edited for clarity.