A groundbreaking study published in JAMA Internal Medicine has unveiled a transformative potential in emergency medical response: an artificial intelligence-driven cardiopulmonary resuscitation (CPR) coaching system known as ChatCPR. Developed through collaboration among leading research institutions including the University of California San Diego, University of Pittsburgh School of Medicine, and Johns Hopkins University, this experimental AI model demonstrates superior performance compared to traditional 911 dispatchers in guiding bystanders through life-saving CPR procedures. This marks a significant leap forward in emergency medicine, integrating advanced AI capabilities in real-time response scenarios where every second can determine survival outcomes.
Cardiac arrest remains a critical public health challenge, with over 350,000 out-of-hospital cases in the United States annually and a dismal survival rate hovering around 9%. Most bystanders lack formal CPR certification, relying heavily on guidance from emergency dispatchers when an incident occurs. The gap between the victim’s collapse and the initiation of effective CPR is crucial; rapid, accurate instruction significantly increases chances of survival. ChatCPR was engineered to address this urgent need by delivering impeccable, guideline-adherent CPR instructions, thereby potentially transforming bystander intervention efficacy during emergencies.
The development of ChatCPR began with a comprehensive benchmarking of existing large language models — including ChatGPT, Claude, Grok, Gemini, Llama, and Mixtral — in simulated emergency scenarios. These models were evaluated against standardized checklists embodying the American Heart Association’s CPR guidelines for diverse real-world situations, such as drownings and sudden cardiac collapses, and for patients across various age groups. While popular AI models demonstrated competency on fundamental CPR steps by scoring around 90% accuracy, their performance suffered on advanced instructions critical to optimizing resuscitation quality, with average scores dropping near 70%.
Recognizing that partial compliance with CPR protocols could translate into fatal consequences, the research team prioritized precision and completeness in AI coaching performance. The genesis of ChatCPR involved iterative refinement grounded in both dispatcher training manuals and clinical resuscitation best practices. This process targeted recurrent failure modes detected in existing models, resulting in an AI agent that achieved a perfect 100% adherence rate on both basic and advanced CPR protocol metrics in controlled simulations, indicating a new standard for AI-guided emergency response.
Beyond simulated evaluations, the study took a crucial step to assess real-world feasibility and efficacy by employing a separate dataset of de-identified 911 call recordings in which human dispatchers had provided CPR instructions. ChatCPR’s instructions were directly compared against those delivered during these actual emergency calls. The results were compelling: ChatCPR outperformed human dispatchers in every metric, scoring 100% adherence to essential CPR steps while dispatchers averaged 85%. On nuanced aspects such as chest compression depth, rate, and chest recoil guidance, ChatCPR’s superiority widened even further, achieving a 99% score in contrast to dispatchers’ 63%.
Experts emphasize that the difference between survival and death in cardiac arrest often hinges on meticulous application of CPR protocols rather than stylistic delivery. ChatCPR excelled in areas historically challenging for human dispatchers, particularly under the duress and cognitive load of managing emergency communications. It ensured consistent, guideline-precise instructions, underscoring AI’s potential to raise the baseline quality of CPR coaching without supplanting human judgment or empathy during crises.
Despite its impressive accuracy, the research team underscores that AI systems like ChatCPR are not infallible and stress that AI integration in emergency medical settings must be accompanied by rigorous, real-world testing for safety and user-friendliness under chaotic conditions. Human oversight remains paramount, with the AI serving as a complementary tool to support but not replace trained professionals. This duality underscores an important paradigm in the evolving role of AI in healthcare—augmenting frontline responders to enhance outcomes while maintaining human decision-making authority.
A distinctive feature of ChatCPR is its open-source availability, inviting developers, researchers, and healthcare organizations worldwide to adopt, adapt, and improve the system. This open science approach aims to accelerate broad adoption and continuous innovation, fostering a collaborative ecosystem focused on refining AI-facilitated emergency care. By providing complete transparency into the algorithms, testing methodologies, and datasets, the authors seek to catalyze cross-platform evolution of life-saving technologies accessible to diverse populations.
Beyond enhancing immediate bystander CPR guidance, the investigators envision AI’s broader integration across the cardiac arrest response continuum. Potential extensions include supporting dispatchers through standardized decision support, assisting first responders and clinicians with scenario-specific training, and enabling adaptive, personalized instructions that respond dynamically to unfolding emergencies. These innovative pathways could collectively drive systemic improvements in cardiac arrest survival rates on a global scale.
Amid the promise of AI-driven resuscitation coaching, ethical and regulatory considerations remain essential to address. Existing legal protections for bystanders performing CPR do not automatically extend to AI-enabled interventions, raising questions about liability, accountability, and informed consent in dynamically mediated emergencies. The study coauthors advocate for development of clear regulatory frameworks to govern AI deployment in medical emergencies, ensuring trust, safety, and equitable access while mitigating risks associated with new technologies.
This pioneering work grounds the exuberant enthusiasm surrounding AI applications in the stark realities of life-and-death scenarios. By rigorously aligning AI capabilities with established medical guidelines and real-world conditions, ChatCPR represents a critical advance toward bridging the fatal interval between cardiac arrest onset and initiation of effective life-saving care. As this technology matures and gains clinical validation, it holds the potential to democratize access to expert-level emergency guidance, empowering bystanders worldwide to act decisively and competently when seconds count most.
The full article, entitled “An Artificial Intelligence–Enabled Cardiopulmonary Resuscitation Instructor,” is accessible through JAMA Internal Medicine’s online platform and includes contributions from renowned experts such as Clifton Callaway, M.D., Ph.D., and Patrick M. Kochanek, M.D., among others. The study signals a transformative inflection point in the fusion of artificial intelligence with clinical medicine, underscoring a future where intelligent systems augment human capacity to save lives in real-time emergencies.
Subject of Research: People
Article Title: An Artificial Intelligence–Enabled Cardiopulmonary Resuscitation Instructor
News Publication Date: 18-May-2026
Web References:
References:
Ayers JW, Desai N, Horvat CM, et al. An Artificial Intelligence–Enabled Cardiopulmonary Resuscitation Instructor. JAMA Intern Med. 2026; doi:10.1001/jamainternmed.2026.1552
Image Credits: Courtesy of John W. Ayers, UC San Diego Qualcomm Institute
Keywords
Artificial Intelligence, CPR coaching, cardiac arrest, emergency medical response, AI safety, 911 dispatch, life-saving technology, open-source AI, resuscitation guidelines, healthcare innovation, AI ethics, real-world AI testing
