Health care is moving through what Pete Clardy, MD, PhD, described as an “epistemic shift” driven by rapidly advancing artificial intelligence (AI). As AI becomes more capable, he described changes in how clinicians process increasingly complex health information, how the technology is implemented and used, how patients engage with that information, and how medicine may need to think about expertise.
Dr. Clardy is a pulmonary and critical care physician who serves as Director of Clinical Enterprise at Google for Health in New York. In this role, he leads a team of clinicians focused on scaling health care and life sciences innovation globally through emerging technologies and research.
Delivering the keynote address at Cleveland Clinic’s AI Summit for Healthcare Professionals, Dr. Clardy traced the evolution of computerized medical decision support from early expert systems and supervised machine learning to large language models, agentic systems, and emerging “world models” that are designed to understand and simulate environments.
Against that backdrop, he characterized the AI technology as still being at an early stage: “AI is in its infancy. It’s as bad as it will ever be right now, and the rate of change is remarkable.”
That rapid evolution is unfolding as clinicians are being asked to make decisions from an ever-expanding volume of information. “We find ourselves collectively in this situation of too much data, not enough information,” he said.
From organizing data to augmenting decisions
Medical training has historically taught clinicians to become strong pattern recognizers, Dr. Clardy explained. But as the “illness presentation” collection of signs and symptoms that clinicians must interpret has become increasingly complex and multimodal, so too has the pattern-recognition task. AI, in Dr. Clardy's view, has a growing role in augmenting clinicians’ ability to make sense of this expanding information environment. That role is also changing.
Earlier machine learning systems in medicine relied on labeled data to learn narrow tasks, such as identifying abnormalities in medical images. Transformer-based systems, by contrast, can learn from data without supervision and generate more complex outputs, with more recent advances extending to agentic frameworks in which multiple agents can perform distinct tasks, as demonstrated by the multiagent AI system, Google’s "Co-Scientist" research.
“What we’re seeing … [is] helping to find signal from the noise, organize and summarize complex information, [but] not take over clinical decision-making,” Dr. Clardy said. “But we are starting to see capabilities that go higher on that developmental pyramid.”
He emphasized, however, that technological sophistication alone does not determine whether organizations successfully implement AI. Rather, successful implementation depends on change management, alignment, identifying the appropriate stakeholders, and being “crystal clear” about the problem to be solved or the goal to be achieved. That includes whether the goal is to automate a process, augment human work, or pursue innovation, he explained, while also encouraging individuals to start with low-risk use cases.
“This is where we fail more often than on the basis of technology,” he said. “We are insufficiently crisp on the problem to be solved.”
Dr. Clardy described the current period as an important “tool shaping moment,” noting that AI tools remain highly adaptable but will “harden over time.” In that context, he emphasized the importance of carefully choosing the goals for their use and invoked Father John Culkin’s observation: “We shape our tools, and therefore our tools shape us.”
Toward an AI co-clinician
Dr. Clardy also described research aimed at progressively expanding what an AI “co-clinician” might be able to do. That work has moved from medical question answering and differential diagnosis toward text-based, back-and-forth interactions as part of triage and intake for an episodic clinical encounter, following a “hill climbing” approach modeled on how medical students are trained to care for patients.
In a single-center study that Dr. Clardy cited, a text-based AI system obtained the history of present illness, incorporated information from the patient’s medical record, and generated clinical documentation and a skeletonized assessment and plan. He said patient trust increased after the AI interaction, whereas AI-generated differential diagnoses and management plans were similar in quality to those generated by humans. The research is now moving into a multicenter study.
Other research has extended the interaction from text to video. Dr. Clardy described a model with two agents working in parallel: a “talker” agent that maintains the conversational flow and a “planner” agent that monitors the video and conversation for areas requiring clarification or discrepancies. He said the AI co-clinician performed near human levels in triage and history taking and outperformed comparison models in clinical reasoning.
He also placed the video-based work in the context of earlier studies of text-based interactions that had identified what he called an “empathy gap,” with AI performing better than humans, whereas humans performed better than video AI for certain types of counseling and communication.
“This is really interesting research and also, yay humans,” he said.
The evolving triadic relationship
For Dr. Clardy, another important trend is how quickly patients are beginning to use AI to understand their own health information, both alongside and independently of clinicians. He characterized the resulting clinician/patient/AI interaction as an evolving “triadic relationship.”
He said these tools may help level the historic information asymmetry between patients and clinicians, with patients able to review their medical records in detail and arrive at visits with more informed questions.
Dr. Clardy connected the rapid adoption of AI by patients to a lesson he took from a National Academy of Medicine discussion about trust and innovation in health care, during which a patient advocate remarked that “innovation in health care moves at the speed of desperation.”
“I think the dichotomy [between] trust and desperation comes when you think about how everyone needs an advocate and a way finder when it comes to managing their own health,” he said in an interview with this news organization. “We in the health-care profession are learning a lot from the ways in which people are using these tools.”
It is, he emphasized during the keynote, “a super important temporal trend to pay attention to.”
Defining expertise in an AI-enabled era
For clinicians, trainees, and medical educators, Dr. Clardy raised an unresolved question about what constitutes medical expertise as AI becomes increasingly capable of memorizing and recalling information, organizing it, and understanding its contextual relevance.
He also highlighted the risks of “deskilling,” “misskilling,” and, for learners who begin their training in an AI-enabled environment, “never skilling”: failing to develop the ability to practice independently in the first place. As he noted, “How we manage [the widespread integration of AI] depends on where we are on our developmental curve.”
What expertise should look like in this environment remains an open question, Dr. Clardy suggested.
“I don’t have an answer,” he said. “But I do think that this is going to be one of the most important questions that we address going forward.”
