An Artificial Intelligence-Enabled Cardiopulmonary Resuscitation Instructor

High-quality, timely cardiopulmonary resuscitation (CPR) after out-of-hospital cardiac arrest (OHCA) is vital, but bystanders need help to act. Telecommunicator-CPR by 911 dispatchers promotes bystander CPR but may be limited by inherent variability and delays in human-led systems.

To evaluate the performance of widely available artificial intelligence (AI) models in delivering guideline-concordant CPR instruction and to develop and assess ChatCPR, a purpose-built AI CPR instructor agent in an early proof-of-concept study design.

In this cross-sectional study, simulated emergency scenarios were used to evaluate baseline CPR instructing capabilities of widely available AI models (ChatGPT, Claude, Gemini, Grok, Llama, and Mistral). Models were evaluated for their adherence to minimally viable criteria, which included instructions like performing appropriate-depth chest compressions, and maximally effective criteria, which included more nuanced instructions like ensuring that compressions achieved full recoil, all derived from major association CPR guidelines. These insights informed development of the purpose-built AI CPR instructor agent that was subsequently evaluated using standard test-retest methods across the same simulated scenarios and 911 dispatcher-assisted calls where CPR was indicated. Data were analyzed from April through December 2025.

In simulated OHCA scenarios among 6 widely available AI models, the models achieved 89.7% (95% CI, 84.8%-93.2%) of minimally viable criteria, ranging from 79.4% (95% CI, 63.2%-89.7%) for Gemini to 97.1% (95% CI, 85.1%-99.5%) for Grok and Claude. Models achieved a mean of 69.8% (95% CI, 65.5%-73.7%) of maximally effective criteria, ranging from 61.3% (95% CI, 50.3%-71.2%) for Llama to 75.0% (95% CI, 64.5%-83.2%) for GPT-4o. The instructor agent achieved 100% (95% CI, 89.8%-100%) and 100% (95% CI, 95.4%-100%) adherence to minimally viable and maximally effective criteria, respectively, for the same scenarios. In a retest using 911 calls, the agent achieved 100% (95% CI, 88.6%-100%) adherence to minimally viable and 98.9% (95% CI, 94.0%-99.8%) adherence to maximally effective criteria, representing absolute improvements of 15.5 percentage points (95% CI, 2.1-27.4 percentage points; P = .02) and 36.1 percentage points (95% CI, 27.2-44.3 percentage points; P < .001) over dispatchers (84.5%; 95% CI, 73.1%-91.6% and 62.8%; 95% CI, 55.0%-70.0%, respectively).

These findings suggest that AI-enabled CPR instruction shows promise for supporting bystanders. Further validation in diverse, general population settings is warranted to define the role of AI-based CPR instruction as a scalable public health intervention for OHCAs.