Patients Already Use the Co-Clinician Doctors Are Testing

Peter Clardy told more than 900 clinicians at Cleveland Clinic on August 28 that medical AI is still in its infancy, and that it will never be this weak again. The Google Health physician spent the keynote on a harder question than the models: what counts as clinical expertise once patients already walk in with a chatbot.

He described an “epistemic shift” in how doctors process information, how patients hunt for it, and what the profession means by skill. The co-clinician Google is testing in simulated visits is one answer. The quieter fact is that a large share of the public did not wait for that answer.

The Second Cleveland Clinic AI Summit

The day-long meeting at the InterContinental Hotel was the clinic’s second AI Summit for Healthcare Professionals, held with the College of Healthcare Information Management Executives. Cleveland Clinic’s newsroom counted more than 900 registered attendees in the room and online.

Clardy, a pulmonary and critical care doctor, is director of the Clinical Enterprise team at Google Health in New York. He consulted with Google on medical records from 2018, kept practicing in the ICU through 2020, and joined the company full time at the end of that year. Before Google he ran the medical intensive care unit at Beth Israel Deaconess Medical Center, directed the Harvard pulmonary and critical care fellowship, and served as a former associate dean for medical education at Harvard Medical School. He also advises the National Academy of Medicine, the Josiah Macy Foundation, and the American Board of Internal Medicine.

Jame Abraham, chairman of hematology and medical oncology at Cleveland Clinic and one of the summit’s organizers, hosted the keynote conversation. “Artificial intelligence has the potential to fundamentally change the way we care for patients, educate future clinicians and advance research,” Abraham said after the session.

Thanks to Dr. Pete Clardy of @Google who delivered an outstanding Keynote lecture @ClevelandClinic Second AI Summit https://t.co/YBrAWo2MJy pic.twitter.com/eARhminE6I

— Jame Abraham, MD, FACP (@jamecancerdoc) August 30, 2026

Clardy’s own line was blunter. “AI is in its infancy. It’s as bad as it will ever be right now, and the rate of change is remarkable,” he said. He traced decision support from early expert systems and labeled-data models to large language models, then to agentic systems and early “world models” built to simulate environments. The clinical problem, as he put it, is older than any of those stacks: “We find ourselves collectively in this situation of too much data, not enough information.”

What an AI Co-Clinician Can Already Do

Medical school still trains doctors as pattern recognizers. Clardy argued that the “illness presentation” those doctors must read is now multimodal and too large for unaided memory. Google’s research answer is AMIE, the Articulate Medical Intelligence Explorer, a system the company still labels as research, not a product in the exam room.

In the 2024 text study, patient actors sat in a chat window with either AMIE or a board-certified primary care doctor. Google Research reported 149 case scenarios drawn from OSCE packs in Canada, the United Kingdom, and India, scored by specialist physicians and by the actors. AMIE posted greater diagnostic accuracy and higher marks on 28 of 32 specialist axes and 24 of 26 patient-actor axes. That design had a built-in tilt. Doctors had to type the way a chatbot types, inside a medium the model was born in.

Google’s Text Study Used a Chat Window

Clardy described a separate single-center intake study in which a text system took a history of present illness, pulled the record, and drafted a note plus a skeleton assessment and plan. He said patient trust rose after the AI pass, and that the machine’s differentials and plans looked similar in quality to the human ones. That work, he said, is moving into a multicenter study. Google has already published a real-world feasibility project with Beth Israel Deaconess Medical Center, Clardy’s old hospital, and an ongoing nationwide randomized study with Included Health.

Video Added a Talker, a Planner, and Eyes

On August 11, Google Research and Google DeepMind put AMIE on a live video link, built on Gemini and Project Astra. Clardy sketched two agents in parallel, a talker that keeps the conversation moving and a planner that watches for gaps. The published system adds a third, a perception agent that reads the audio and video stream for distress, gait, breathing, and other cues a pause would miss.

The video OSCE used 100 clinical scenarios across five body systems. Fifteen trained patient actors completed 300 standardized visits in three arms: AMIE on video, AMIE in text, and 10 board-certified primary care physicians on the same video interface. A separate panel of 20 experienced primary care doctors scored the visits. Google reported expert-level performance in real-time video on history, diagnosis, management, and communication, with a clear edge on eliciting physical signs and walking actors through virtual exam maneuvers. The actors preferred video to text and rated AMIE well on empathy and rapport.

AMIE AGAINST PRIMARY CARE IN SIMULATED VISITS

Google states the limit in plain language. Nobody in those 300 visits was a real patient with an unscripted complaint. Occasional perception and reasoning errors still show up in automated tests, and the Astra prototype still drops conversational smoothness. Promising and proven in clinic are not the same sentence, and the video paper does not pretend they are.

The same lab family now includes Co-Scientist, a multi-agent system that debates hypotheses in a tournament and then offers them to working scientists. In one Stanford collaboration, a repurposing candidate blocked 91 percent of a scarring-linked response in liver-fibrosis lab tests. That is a research partner for investigators, not a bedside doctor, and Google says so.

A Third of Adults Already Use Health Chatbots

Clardy spent part of the keynote on a triad the profession still treats as a side plot: clinician, patient, and AI, with the patient often arriving first. Pew Research Center surveyed 3,488 U.S. adults from June 22 to 28, 2026, and found that 34 percent of Americans use health chatbots for at least one of eight medical reasons. That is not a future scenario. It is last summer’s behavior.

WHY AMERICANS OPEN A HEALTH CHATBOT

• Speed: 28 percent want health information quickly.
• Symptoms: 25 percent try to figure out what is causing them.
• Cost: 22 percent want the information at little or no cost.
• Treatments: 22 percent ask what a treatment involves.
• The doctor’s words: 22 percent go back to learn more about a diagnosis they already heard.
• Lab results: 20 percent ask a bot to explain the numbers.
• Awkward topics: 18 percent raise things they do not want to say out loud.
• The decision to go: 15 percent use a bot to decide whether to book a visit at all.

Among those chatbot health users, 47 percent called the information extremely or very helpful and 48 percent called it somewhat helpful. Only 5 percent said it was not. Use skews young: 44 percent of adults 18 to 29 qualify as chatbot health users, against 17 percent of adults 65 and older. In a companion cut of the same survey, 72 percent of U.S. adults said it is extremely or very important that a doctor tell them if AI is in their care.

Younger patients are already further along. A 2025 JAMA Pediatrics survey found 19.2 percent of adolescents and young adults had used a chatbot for mental-health advice, and 63.3 percent of those users had told no one. Most still rated the advice as somewhat or very helpful. The visit, when it happens, starts after that private conversation, not before it.

Clardy tied the rush to a line he heard from a patient advocate at a National Academy of Medicine discussion on trust and innovation: “innovation in health care moves at the speed of desperation.” He treated that as a temporal trend, not a slogan.

I think the dichotomy between trust and desperation comes when you think about how everyone needs an advocate and a wayfinder when it comes to managing their own health.

Peter Clardy, MD, director of the Clinical Enterprise team, Google Health, in an interview after the Cleveland Clinic keynote

The information gap that used to run one way is now a three-body problem. Patients can reread their records, query a model at 1 a.m., and arrive with a differential they did not get from the office. Doctors who ignore that prep will spend the first minutes of the visit catching up to a machine the patient already trusts for speed.

Trainees Face a Skill They May Never Form

For the people still in training, Clardy named three failure modes that depend on where you sit on the developmental curve. Experienced clinicians can lose a skill they once had. They can also learn the wrong lesson from a fluent, wrong model. Trainees can fail to build the skill at all.

THE THREE SKILL RISKS

• Deskilling: A doctor who used to do the work unaided gets rusty once the model does it first.
• Misskilling: A fluent error lands as fact, and the learner stores it as knowledge.
• Never-skilling: A student never builds independent judgment because the answer arrived before the struggle that forms it.

The colonoscopy evidence is the sharpest empirical warning we have so far, and it is about people who already knew the skill. A 2025 multicenter observational study led by Krzysztof Budzyń in Lancet Gastroenterology and Hepatology found adenoma detection among experienced endoscopists fell from about 28 percent to about 22 percent in the months after AI support was withdrawn. Tyler Berzin, writing in The Lancet that October, treated that drop as a reason to protect core skills on purpose, not as a reason to shelve the tools.

Yuhe Ke and colleagues, in a May 2026 Nature Medicine perspective, argued that never-skilling is the under-discussed risk for students and residents. Direct trials inside medical school are still thin. The learning-theory case is not. If a model writes the differential before the intern has to retrieve an illness script, the retrieval practice that builds expertise never happens. Ke’s group offered a three-phase guardrail: prove AI-independent baseline competence first, teach calibration on purpose, then let trainees use models under supervision.

Clardy did not pretend he had the curriculum. “How we manage the widespread integration of AI depends on where we are on our developmental curve,” he said. On what expertise should look like once recall and organization are cheap, he was frank. “I don’t have an answer.”

The Failure Mode Is Fuzzy Goals

The part of the talk least likely to make a demo reel was the implementation warning. Clardy said success depends on change management, stakeholder alignment, and being “crystal clear” about whether a project is meant to automate a task, augment a person, or chase a new idea. Start with low-risk uses, he said. “This is where we fail more often than on the basis of technology. We are insufficiently crisp on the problem to be solved.”

That is a more useful sentence than most model cards. A hospital can buy a strong summarizer and still drown attendings in notes so long that another model has to summarize the summary. The loop is already visible in enterprise rollouts: one system generates dense text, a second system compresses it, and the human in the middle is asked to sign both. If the goal was less clerical load, the stack missed.

Google’s own papers keep repeating a related limit. AMIE is scored in OSCEs, with actors, on conditions that can be performed on camera. It is not a license to let a model close a clinic. Clardy put the current job lower on the pyramid: find signal in noise, organize, summarize, and leave the decision with the clinician, even as the research climbs toward triage, intake, and video presence.

Clardy Called This a Tool Shaping Moment

He named the period a “tool shaping moment.” The systems are still soft. They will “harden over time.” Father John Culkin’s line, which Clardy invoked, is the whole policy problem in one clause: “We shape our tools, and therefore our tools shape us.”

AI is in its infancy. It’s as bad as it will ever be right now, and the rate of change is remarkable.

Peter Clardy, MD, keynote, Cleveland Clinic AI Summit for Healthcare Professionals, August 28, 2026

If the tools harden around actor-tested co-clinicians while a third of adults already treat a consumer chatbot as a wayfinder, the profession will have defined expertise after the fact. The people with the most skin in that delay are not the labs publishing OSCEs. They are the patient who pastes lab results into a prompt at night, and the intern who never had to build a differential without one.

Clardy closed the expertise question the way a careful educator does, without a slogan. He does not have the answer. He thinks it is one of the questions the field will have to answer while the clay is still wet.

Disclaimer: This article is news reporting and analysis of public research and a conference keynote. It is informational only and is not medical advice, a diagnosis, or a treatment recommendation, and it does not tell readers to use or avoid any chatbot, app, or clinical AI system for their own care. Anyone with symptoms, test results, or treatment questions should talk with a licensed physician or other qualified clinician who knows their history before acting on information from a model or a news story. Study figures, product statuses, and trial designs reflect the papers and statements cited here as of the dates on those sources and can change as new evidence is published.