Editor's Note 

AI and Clinical Care

Artificial Intelligence and Bystander Cardiopulmonary Resuscitation—Pushing Forward

Teva D. Brender, Sharon K. Inouye, Cary P. Gross

JAMA Intern Med Published Online: May 18, 2026

doi: 10.1001/jamainternmed.2026.1559

Early administration of high-quality cardiopulmonary resuscitation (CPR) and defibrillation are the most effective interventions for out-of-hospital cardiac arrest (OHCA).¹ Yet bystander CPR occurs in only half of OHCA cases,¹with lack of knowledge and training commonly cited as key barriers.² Bridging this gap is a central challenge in resuscitation science.

In this issue of JAMA Internal Medicine, Desai et al³ explore whether large language models (LLMs) can support bystander CPR. First, they evaluated 6 general purpose, commercially available LLMs on 5 simulated emergency scenarios. The models averaged 90% adherence to a core set of quality metrics (range across models, 79%-97%) and 70% adherence to a comprehensive quality checklist (range, 61%-75%). Building on these findings, the research team developed ChatCPR, a bespoke CPR instructor using an open source LLM.³ When 911 calls from an open access repository were transcribed and provided in a text format, the LLM CPR instructor achieved nearly perfect (99%) comprehensive checklist adherence and outperformed human dispatchers (63%).³

Given these findings, should emergency dispatchers start looking for other jobs? No, far from it. The LLM tool, and others that are sure to follow, will need to be trained and evaluated on a broader, higher-fidelity set of clinical patient scenarios. For instance, CPR was indicated in 4 of the 5 simulated and all of the emergency scenarios studied.³ An artificial intelligence (AI) CPR instructor must reliably recognize when CPR should not be performed. Furthermore, this study assessed text-based LLMs. To be practical for clinical use, an AI CPR instructor needs speech-to-speech functionality (processing audio input and generating natural speech output) and be able to handle multiple languages, ambient noise, and situations with limited connectivity.

For the AI CPR tool to be effective, it would need to integrate efficiently and seamlessly into the workflow of the emergency dispatcher, which would require extensive logistical testing and evaluation. OHCA represents only a small share of all emergency calls (eg, <2% [12 of 743] in the public repository).⁴ Emergency dispatchers are conducting manifold other functions, including triage and management of a range of social and medical emergencies, encompassing sometimes complex situations. such as mass shootings, domestic violence, missing children, drug overdoses, and mental health crises. Use of an AI CPR instructor in actual clinical care should focus on strategies that enhance and augment the role of the emergency dispatcher, the essential “human in the loop,” by helping with performance of routine, structured tasks. This approach would enable dispatchers to provide the highest-quality recommendations for protocolized care, such as CPR, while retaining control over complex decision-making that incorporates contextual understanding, judgment under uncertainty, ethical reasoning, creativity, and flexibility.

In this proof-of-concept study, Desai and colleagues³ push the field forward by demonstrating that an AI tool can deliver high-quality, guideline-concordant CPR instructions in the targeted situation of transcribed 911 calls. Future developers and researchers should build on their transparent, open source design to independently assess, modify, and improve the AI CPR tool; test model performance on larger, more diverse datasets with clinical considerations; study different emergency response implementation strategies that integrate seamlessly with the role of the emergency dispatcher; and, eventually, evaluate when and how an AI CPR instructor can enhance (and definitely not replace) our current human-based approach to advising callers on bystander CPR and improve clinical OHCA outcomes.