An AI-powered CPR coaching agent, ChatCPR, outperformed 911 dispatchers in real-time guidance, meeting all guideline-based basic steps and 99% of advanced steps versus 85% and 63% for human dispatchers, respectively.

Researchers benchmarked major AI models and found ChatCPR uniquely designed to handle advanced, guideline-critical details beyond basic guidance provided by other models like ChatGPT, Claude, and Gemini.

ChatCPR was developed by UC San Diego with collaborators from Johns Hopkins and UPMC to guide bystanders through CPR during emergencies, aiming to improve survival in out-of-hospital cardiac arrests.

The project is open-source to foster global adoption, adaptation, and continuous innovation, with the goal of democratizing access to expert-level emergency guidance.

Researchers stress ongoing real-world testing, safety, and user-friendliness amid chaotic conditions, and call for clear regulatory frameworks around liability and accountability in AI-assisted emergencies.

Future enhancements envision on-device operation and broader accessibility to prevent CPR guidance from becoming luxuriously scarce in resource-limited regions.

Broader integrations are envisioned, including dispatcher decision-support, training aids for first responders, and adaptive, personalized instructions as emergencies unfold.

The effort is a collaboration among UC San Diego, the University of Pittsburgh School of Medicine, and Johns Hopkins, centered on improving emergency response where every second counts.

Implementation remains the primary challenge for healthcare AI, and open-sourcing is expected to accelerate adoption and reduce disparities by broadening access to CPR guidance.

ChatCPR has been released freely for use and study to encourage further refinement and deployment across platforms.

ChatCPR is presented as a complementary tool to human responders, with safeguards, real-world testing, and integration with 911 systems emphasized.

The study advocates for clear regulatory frameworks and legal protections for bystander intervention with AI, highlighting the need for ongoing human oversight and policy development.