Our Take
ChatCPR is the rare healthcare AI claim backed by head-to-head testing against an actual human baseline (911 dispatchers on real calls), not marketing benchmarks—but open-sourcing it means adoption speed, not immediate lives saved, is the real test.
Why it matters
Nearly 90% of out-of-hospital cardiac arrests in the US are fatal, yet only 2% of Americans are CPR-certified. Every minute without CPR reduces survival odds. If ChatCPR reaches phones as a tool that works offline, access becomes the lever—not model sophistication.
Do this week
Emergency services directors: request the open-source ChatCPR code and test it on your dispatcher call logs this quarter so you can measure whether it closes knowledge gaps in your own call center before committing to deployment.
UC San Diego, Johns Hopkins, and UPMC Release Open-Source CPR AI
Researchers led by John Ayers at UC San Diego's Altman Clinical and Translational Research Institute published ChatCPR in JAMA this week. The tool is an AI agent that coaches bystanders through CPR in real time.
In testing against recordings from actual 911 calls, ChatCPR outperformed human 911 dispatchers at guiding callers through CPR. The researchers benchmarked major models—ChatGPT, Claude, and Gemini—on CPR instruction tasks. Off-the-shelf models performed reasonably on basic guidance but missed clinically important nuances. ChatCPR was built on a smaller, lower-performing language model but engineered specifically for CPR guidance with domain-specific training and careful prompt design.
The team released ChatCPR as open-source rather than a commercial product. They made training materials, guidelines, prompts, and architecture publicly available so emergency-response organizations and companies can build on it, improve it, and deploy it.
The Real Test Is Implementation, Not Model Power
Ayers has been vocal about healthcare AI hype. His 2023 JAMA study found that AI chatbots write more accurate and empathic responses to patient messages than human doctors—but he has emphasized that accuracy on a test set does not mean lives saved. After a reporter asked him on-air whether AI would save lives, Ayers realized the gap between demonstration and impact. "It's not really helping anybody yet—it's all hype, no reality," he said.
ChatCPR targets a specific, high-stakes problem where seconds matter. More than 350,000 Americans suffer cardiac arrests outside hospitals each year. Only 2% are CPR-certified. Most people call 911 and wait. Every minute without CPR reduces its efficacy.
Ayers noted that the key challenge in healthcare AI is implementation. ChatCPR was intentionally built on a small model that could eventually run directly on smartphones without internet connectivity. That design choice—favoring offline availability over raw model capability—signals a different priority: availability beats sophistication when you need help in the moment.
Open-sourcing the tool also addresses equity. CPR coaching should not depend on where you live or what emergency services budget your area has. Broad deployment depends on adoption by multiple vendors and agencies, not proprietary control.
Deployment Is the Bottleneck, Not the Science
The JAMA study provides independent validation (peer-reviewed testing against real 911 dispatcher calls), not vendor marketing. That is rare in healthcare AI. But validation in a study does not mean ChatCPR will reach the people who need it.
Emergency services organizations should treat this as a tested prototype, not a finished product. The next phase is integration: Can dispatchers or callers use it on phones they already carry? Does it work offline? Does it integrate with existing CAD (computer-aided dispatch) systems? Does retraining improve outcomes further?
The open-source model means fragmented adoption is likely. Some regions may fork it, tune it locally, and never share improvements back. Others may license commercial builds on top of it. The risk is that ChatCPR becomes a reference implementation that sits in GitHub while bystanders still call 911 unprepared. The win is that any jurisdiction with engineers can implement it without waiting for a vendor's commercial roadmap.