A new study from scientists at the University of California San Diego in collaboration with the University of Pittsburgh School of Medicine, Johns Hopkins University, and other institutions, demonstrates that an artificial intelligence-powered CPR coaching agent can outperform 911 dispatchers in guiding bystanders through cardiopulmonary resuscitation.
The study, published in JAMA Internal Medicine , first benchmarks popular AI models on CPR instruction and then introduces ChatCPR, an open-source AI agent that scored 100% on guideline-based CPR checklists and significantly outperformed human dispatchers when tested against recordings from previous real 911 calls.
“If AI is going to earn its place in medicine, it should start by helping people save the person right in front of them,” said John W. Ayers , Ph.D., who is a UC San Diego Qualcomm Institute scientist, in addition to Head of AI at the UC San Diego Altman Clinical and Translational Research Institute, professor at the UC San Diego School of Medicine, and study coauthor.
“More than 350,000 Americans suffer out-of-hospital cardiac arrest each year, and survival sits at roughly 9%. Given that only 2% of Americans are certified to perform CPR, when someone collapses, they call 911 and wait. ChatCPR could change that and begin to save lives,” added Nimit Desai, B.S., a research affiliate at the UC San Diego Qualcomm Institute, a medical student at UC San Diego School of Medicine, and lead author of the study.
Adapting new technologies could help close the gap between cardiac arrest and initiation of CPR. In this proof-of-concept study, researchers used 12 real 911 calls to compare AI-generated CPR instructions with those provided by dispatchers.
“As new AI technologies emerge, we know people are going to start using them in real-world situations,” said Christopher M. Horvat , M.D., Director of Medical Emergency Response Teams at UPMC Children’s Hospital of Pittsburgh. “Our goal was to take a first step to understand how these tools perform and how they should be evaluated before being used in patient-facing settings,” said Horvat, who is also Associate Director of the Safar Center for Resuscitation Research at the University of Pittsburgh, named for Dr. Peter Safar, a pioneer of modern CPR.
The findings highlight an opportunity to study AI’s role in delivering guideline-based instructions—as a complement, not a replacement, for human responders. “This is about supporting people in high-stakes situations where human judgment is essential,” added Horvat. “The goal is to raise the floor of performance, not to replace trained professionals.”
The research team first benchmarked popular AI models, including ChatGPT, Claude, Grok, Gemini, Llama, and Mixtral, on CPR coaching in simulated emergency scenarios, evaluating them against a checklist of criteria for delivering guideline-concordant CPR instruction in out-of-hospital cardiac arrest.
Across different situations, like drowning or collapsing while jogging, and across different patients, from toddlers to seniors, AI models performed well on the basics of CPR. On average, they scored 90% on essential steps, such as where to press on the chest and how fast to do it. Scores ranged from 79% (Gemini) to 97% (Grok and Claude). When it came to giving the best possible instructions to improve survival, performance dropped. On these more advanced steps, such as letting the chest fully rise between compressions, models averaged 70%. Scores ranged from 61% (Llama) to 75% (ChatGPT).
“In cardiac arrest, good is not good enough,” said Cameron Dezfulian, M.D, an adult and pediatric intensivist, senior faculty member at Baylor College of Medicine, and study coauthor. “Missing 10 to 30% of steps can be the difference between life and death.”
Those gaps informed the development of ChatCPR, an open-source AI agent for CPR coaching grounded in 911 dispatcher training materials and CPR best practices. The system was iteratively engineered to address specific failure modes. In the same scenarios, ChatCPR scored 100% on simulated calls on both the basic steps and the more advanced steps needed to give someone the best chance for survival.
The critical question was whether it could work in real life. The team used a separate set of real, de-identified 911 calls that were publicly available. In these calls, dispatchers had already provided CPR instruction. The team then compared the dispatchers’ instructions to ChatCPR’s instructions.
“ChatCPR won every head-to-head comparison with human dispatchers,” explained Noor Majhail, B.S., an EMS responder and study coauthor. ChatCPR scored 15 percentage points higher than the dispatcher on basic CPR steps; specifically, dispatchers met 85% of the guideline steps, and ChatCPR met 100%. For more advanced steps, the gap was even larger; ChatCPR scored 99% while dispatchers scored 63% — a gap of 36 percentage points.
“ChatCPR excelled in patient assessment, chest compression depth and rate instructions, and recoil guidance areas where stressed, multitasking dispatchers most often faltered,” added Desai.
“This wasn’t about style. It was about strict adherence to CPR guidelines where precision matters most,” said Davey Smith, M.D., Professor at the UC San Diego School of Medicine, Director of the Altman Clinical and Translational Research Institute at UC San Diego and study coauthor. “ChatCPR addressed elements that dispatchers, under the stress and complexity of real calls, sometimes missed, presenting a valuable opportunity to translate AI into real-world healthcare practice.”
"No AI system is perfect, but the triangulation of problem identification, deep substantive expertise and AI is the crossroads for healthcare to unlock meaningful breakthroughs," said Rema Padman, Ph.D., Trustees Professor of Management Science and Healthcare Informatics in the Heinz College of Information Systems and Public Policy at Carnegie Mellon University and study coauthor.
The researchers say careful, real-world testing is still needed. They want to ensure the system is safe, works in chaotic settings, and is easy for people to follow. They also stress the importance of safeguards and the integration with existing 911 systems and continued human oversight.
Towards that goal, the team made ChatCPR open and free for anyone to use and study. They shared the complete system, how they tested it, and all related materials. “Any developer or organization can freely use, adapt, and deploy ChatCPR,” noted Mark Dredze, Ph.D., a Professor of Computer Science at Johns Hopkins and study coauthor. “We encourage researchers to refine, test, and improve this technology across all platforms to provide life-saving assistance for all.”
“AI could add value across the entire cardiac arrest response continuum, helping bystanders start CPR sooner, support dispatchers with standardized guidance, and assist clinicians and first responders with complex or scenario-specific instructions during training,” added Horvat.
Clear regulatory frameworks, as some of the author team has previously noted in an earlier JAMA piece , will also be essential as the tool moves from research to real-world use. “Today, bystanders have strong legal protections against civil and criminal prosecution when intervening to perform CPR, added Mike Hogarth, M.D., Professor of Medicine and Director of Informatics at the Altman Clinical and Translational Research Institute at UC San Diego and study coauthor. “How these protections extend to AI-enabled CPR is a challenge that needs to be addressed.”
“Ultimately, our work grounds AI hype in life-or-death reality,” concluded Ayers. “The real promise is closing the deadly gap between a person collapsing and lifesaving care beginning.”
The article, “An Artificial Intelligence–Enabled Cardiopulmonary Resuscitation Instructor” (doi:10.1001/jamainternmed.2026.1552) included Clifton Callaway, M.D., Ph.D. and Patrick M. Kochanek, M.D. (UPMC & University of Pittsburgh) as additional coauthors. It is available online on the JAMA Internal Medicine website.
JAMA Internal Medicine
10.1001/jamainternmed.2026.1552
Experimental study
People
An Artificial Intelligence–Enabled Cardiopulmonary Resuscitation Instructor
18-May-2026
Dr Desai reported receiving consulting fees from Pearl Health outside the submitted work. Dr Dredze reported receiving personal fees from Bloomberg LP, Good Analytics, and Medeloop outside the submitted work. Dr Hogarth reported receiving personal fees from LifeLink outside the submitted work. Dr Smith reported serving as a consultant for Model Medicines, Hyundai Biosciences, Capricor, Gilead, and Pfizer. Dr Callaway reported being a cofounder of Intellicardio, Inc outside the submitted work. Dr Ayers reported receiving personal fees and stock from Good Analytics, HealthWatcher, and Medeloop and building artificial intelligence analytical agents for health systems and life science companies outside the submitted work.