Can AI really determine if a caller is in cardiac arrest?

Systems like Corti are designed to analyze ongoing calls, identify key indicators, and alert dispatchers, a technology that agencies like the Seattle Fire Department have used for years. However, its role is strictly assistive—it increases the likelihood of early detection, but final judgment and instructions remain the responsibility of dispatchers, not direct AI-driven decisions.

Will Taiwan's 119 emergency services adopt these systems?

Unlikely in the short term. Taiwan's emergency dispatch architecture differs from the US NG911 standard; the data fields, incident classifications, and regulatory premises of these systems are built around the US framework, meaning localization would require essentially rebuilding them from scratch. What is more likely to appear first is automated triage for non-emergency citizen hotlines, which involves much lower barriers and risks.

Why are gunshot detection systems so controversial?

The controversy centers on three points: false positive rates, cost-effectiveness, and whether they lead to over-deployment in specific communities. For years, independent studies and municipal audits have questioned their return on investment, leading cities like Chicago to terminate or scale back their contracts. In the US, this is already a political issue, not merely a technical evaluation.

What is the biggest risk of these systems?

Data governance. Centralizing previously fragmented police data and making it easily searchable simultaneously increases operational efficiency and the potential for abuse. Retention periods, query permissions, and audit trails must be designed before deployment; trying to patch them in afterward rarely works.

Multiple emergency dispatch centers in the US are already deploying AI to field non-emergency calls, transcribe reports in real-time, and detect cardiac arrest even before the caller finishes speaking. While the efficiency gains of this shift are alluring, the controversies are just as significant—and what Taiwan needs to learn may not be the technology itself.

When AI Starts Answering 911: The Technical Transition Behind Emergency Calls

At 2:40 AM, a call hits the dispatch center. The caller speaks frantically while others shout in the background. The dispatcher tries to calm them down while attempting to extract the address. Forty seconds into the call, a prompt pops up on the dispatcher’s screen: Suspected cardiac arrest, recommend immediate CPR instruction.

This assessment wasn't made by the dispatcher. It was made by a model listening to the same call.

Such systems have been operating in parts of dispatch centers in the US and Denmark for a few years now. They simultaneously showcase the most convincing side of AI entering public services—and the side that requires the most rigorous oversight.

Background

The predicament of emergency dispatch centers is universal: staff shortages, high turnover, long training periods, and zero room for error on every single call. Dispatch centers across multiple US counties have faced severe staff shortages in recent years, leading to longer call wait times—and in scenarios like cardiac arrest where survival rates are measured in minutes, waiting time equals life or death.

Consequently, this traditionally ultra-conservative field has become an unexpectedly active arena for AI adoption. The technological approaches generally fall into four categories.

The first is real-time call interpretation. Corti's machine learning platform analyzes ongoing calls, identifies critical conditions like cardiac arrest, and alerts dispatchers. The Seattle Fire Department is among the early adopters.

The second is non-emergency call routing. A significant proportion of calls received by dispatch centers don't actually require immediate dispatch—noise complaints, illegally parked vehicles, and process inquiries. Aurelian specializes in automating this segment so that true emergency calls don't have to wait in line. This Seattle-based startup closed a $14 million Series A funding round in August 2025.

The third is the rewriting of entire intake and dispatch platforms. Carbyne focuses on Next Generation 911 (NG911), while Mark43 integrates dispatch, records, mobile units, and analytics into a single cloud-based system. Coupled with ReportAI for auto-generating report drafts and BriefAI for summarizing shift briefings, the company reports adoption by over 300 agencies.

The fourth is cross-system data integration. This category is the most unassuming yet most critical: Peregrine connects a dozen disconnected legacy police and fire systems, cleaning and aligning data so that duty officers can review complete contexts in one place. The company recently closed a $250 million Series D funding round.

Industry consolidation is happening at a notable pace. In 2025, Axon acquired Prepared—another 911 AI firm—and acquired Carbyne at a $625 million valuation. This sector is shifting rapidly from a chaotic startup battleground to an oligopoly.

Key Takeaways

AI's role is auxiliary, not decisive: Current systems are designed to "alert dispatchers," with ultimate judgment remaining in human hands. This line must be held in high-risk domains.
Time saved is primarily administrative: Law enforcement and firefighters complain most about filing reports before heading home. Features like ReportAI and BriefAI target this exact pain point rather than replacing frontline judgment.
Data integration is harder than models: Peregrine adopts an implementation model featuring engineers stationed on-site at client locations because public sector data formats are too chaotic for remote processing. There are no shortcuts here.
Controversies are equally real: ShotSpotter gunshoting detection by SoundThinking has faced long-standing scrutiny regarding false-positive rates and cost-effectiveness, leading cities like Chicago to terminate or scale back contracts. When algorithms start deciding where police forces deploy, societal debate is essential.

Market Impact Analysis

For users in Taiwan: The impact will likely be limited in the short term. Taiwan's 119 and 110 systems differ from the US architecture and lack the NG911 specification. The service most likely to see early implementation is "non-emergency case routing"—automating citizen hotlines like Taipei's 1999 service, which carries a much lower technical barrier and minimal risk.

For enterprise applications: The spillover effect of this technology trend is worth noting. Urgent needs for hospital emergency rooms, large venues, and campus entrance security are genuine. Walk-through weapon detection systems like those from Athena Security and license plate/incident recognition from Rekor are expanding into medical and commercial sectors. Hospital violence is not uncommon in Taiwan, creating a real demand here.

For developers: The true opportunity lies at the integration layer rather than the model layer. The degree of fragmentation in Taiwan's public sector systems is on par with the US. Whoever can connect existing data, align semantics, and provide intuitive query interfaces will hold immense value. The barrier to entry is high, but user stickiness is equally strong.

Future Development Trends

I anticipate three main directions over the next two to three years:

First, shifting from alerts to draft generation. AI will not only prompt "this might be a cardiac arrest," but will also draft the entire case record, leaving humans to simply review and confirm. Acceptance of document automation is far higher than decision automation and will become widespread first.

Second, regulators will catch up. The EU AI Act has classified law enforcement use cases as high-risk, and various US states are successively legislating retention periods for license plate recognition data. The phase where technology races ahead while regulations lag behind is coming to an end.

Third, integration and acquisitions will continue. Axon's consecutive acquisitions of Prepared and Carbyne are just the beginning. Public sector procurement favors single-vendor windows, a structural dynamic that will force suppliers to merge into massive platforms.

TheAI Academy Summary & Review

While writing this piece, what troubled me most wasn't how impressive the technology is, but rather that the performance metrics for these systems are almost entirely self-reported by vendors. Claims such as "response times reduced by double-digit percentages" or "severe crime dropped by 45%" require independent verification at the public policy level, which is largely missing today. ShotSpotter's controversy stems precisely from this: it's not that the technology is completely useless, but that the claimed benefits fail to justify the input costs and social price tag.

Review: When AI enters public safety, the real questions to ask are not "how accurate is it?" but rather "who verifies it, how long is data retained, who can query it, and are queries logged?"—technical problems have solutions, but governance problems have no standard answers.

Specific advice for Taiwanese readers: If you work in the public sector or related supply chains, the most worthwhile investment right now isn't rushing to deploy models, but getting your underlying data in order—field definitions, code mappings, and retention policies. The US experience clearly shows that agencies with messy foundational data see their AI projects fizzle out into nothingness. If you are a general citizen, what deserves your attention is forthcoming personal data regulations: once license plate recognition and image analysis become widespread, "how long data is retained and who has the right to query it" will directly impact our daily lives. This is a critical window to make your voice heard before policies solidify. To explore related tools, check out the site's AI Public Safety and Incident Response Category.

Sources

Mark43 Official Website (Product lines, number of adopting agencies)
Peregrine Official Website (Platform positioning, implementation model)
SoundThinking Official Website (ShotSpotter and other product descriptions)

Compiled based on public information; official sources prevail. Performance figures disclosed by vendors in the article have not been independently verified by a third party. Please exercise discretion when citing.

When AI Starts Answering 911: The Tech Shift Happening Behind Emergency Calls