From the Emergency Services edition of September 30, 2026
UC San Diego: ChatCPR open-source AI beats 911 dispatchers on CPR guideline checklists in proof-of-concept
UC San Diego Qualcomm Institute / University of Pittsburgh / Johns Hopkins et al. (JAMA Internal Medicine study) · Academic EMS / OHCA telecommunicator-CPR research · United States (multi-institution; tested on public U.S. 911 call recordings)
- Publisher
- UC San Diego Today
- Original publication
- 2026-05-18
- Source retrieved
- 2026-10-01
- Event date
- 2026-05-18
What happened
May 18, 2026 UC San Diego Today summarizes a JAMA Internal Medicine proof-of-concept: ChatCPR, an open-source AI CPR instructor (built on Llama 3.3 70B Instruct Turbo per related coverage; grounded in T-CPR materials), scored 100% on simulated OHCA checklist criteria and, on 12 real de-identified 911 CPR calls, 100% minimally viable and ~99% maximally effective criteria versus human dispatchers at 85% and 63% (about +15 and +36 percentage points). Authors stress support-not-replace framing, need for real-world trials/safeguards/legal clarity on AI-enabled CPR protections, and open release for adaptation. Not a prospective live-dispatch outcome study.
Why it matters
Measured OHCA instructional-AI evidence for 911/EMS alongside today’s Lake County EMD QA procurement—shows life-safety coaching AI advancing while still pre-operational. Why now: not archived; newly relevant as PSAPs scale assistive AI and need checklist-grade T-CPR evidence with explicit real-world-validation caveats after 2026-09-28.
Evidence and measured results
Opened UCSD Today HTML via curl. Date May 18, 2026; JAMA Internal Medicine citation; 100%/~99% vs ~85%/63% dispatcher comparisons; 12 real 911 calls; support-not-replace and real-world-trial caveats; open-source release statements from page.
Limitations and uncertainty
Proof-of-concept on 12 public calls and simulations; checklist adherence ≠ survival benefit; institutional news release summarizing the journal article.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-09-30; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Telecommunicator-CPR quality is inconsistent under stress while OHCA survival remains low.
- Stakeholders
- EMS medical directors
- ECC training leads
- Telecommunicators
- Legal/risk
- Discovery questions
- Current T-CPR audit scores?
- Sim lab available?
- Liability stance on AI CPR prompts?
- Appetite for open-source evaluation?
- Value hypothesis
- Open CPR agents may standardize high-quality prompts if prospectively validated—checklist deltas on 12 calls are not field proof.
- Potential engagement
- Medical-director-led simulation comparison before any live assist pilot.
- Unsupported claims
- Survival benefit proven
- Replaces 911 dispatchers
- Ready for unsupervised caller use
Pre-sales engineering
Role takeaway
- Fit
- Protocol-grounded CPR instruction agent for training or HITL dispatcher assist.
- Architecture and integration
- Call/sim audio → ASR/dialogue agent → prompt cards to telecommunicator or trainee → human-spoken guidance.
- Prerequisites
- Digitized T-CPR protocol
- Override/barge-in controls
- Logging
- Data-use approval
- Constraints
- Small real-call n
- Retrospective design
- No live outcome data
- Security
- Strict handling of 911 training audio; disable vendor secondary training use.
- Proof of value
- Prospective protocol adherence
- Time-to-first compression
- Dispatcher accept/edit rate
Delivery
Role takeaway
- Work
- Stand up sim evaluation, train users, and draft prospective HITL protocol with legal review.
- Dependencies
- Medical direction
- ECC training
- Counsel/risk
- Research collaborators
- Ownership
- EMS medical director owns clinical content; ECC training owns rollout.
- Skills and adoption
- Teach dispatchers to use agent prompts as checklists they still voice and adapt.
- Governance checkpoints
- IRB/data gate
- Sim go/no-go
- Pre-live liability review
- Acceptance criteria
- Human always speaks to caller
- Prospective trial registered
- Failover without agent
- Risks
- Bystander confusion
- Deskilling
- Overgeneralizing from n=12
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
Position as dispatcher-assist or training/sim tool with human telecommunicator control of DA-CPR—not autonomous caller-facing replacement without prospective trials.
Governance
Who approves, reviews and stays accountable for outcomes?
Require prospective field trials, Good Samaritan/liability clarity for AI-guided CPR, and medical-director approval before any live prompt use.
Security and privacy
What data, permissions and controls need testing?
Retrospective use of real 911 call recordings needs lawful de-identification regimes; live assist would process highly sensitive audio—minimize retention.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
May raise the floor of T-CPR instruction quality if integrated as prompts for multitasking dispatchers; training must prevent automation complacency.
Procurement
What should contracts, pricing and exit terms secure?
Prefer open evaluation materials over closed T-CPR copilots; budget prospective HITL trials rather than buying on checklist deltas alone.
Operating model
Which teams own the service once it runs?
Keep telecommunicators responsible for DA-CPR delivery; use ChatCPR-like agents for standardized prompts and training QA.
Publication history
- 2026-09-30Emergency Services · Issue 245 resources
Stable resource ID: ucsd-chatcpr-ohca-911-dispatcher-benchmark-20260518