Emergency call centers across the United States are testing artificial intelligence systems designed to sort non-emergency calls from urgent ones, freeing human dispatchers to focus on life-threatening situations. At least three separate efforts, spanning a California sheriff’s department, a Nashville emergency communications center, and a Seattle 911 operation, have produced early data on how AI can reduce wait times and improve triage accuracy. The results so far are promising but incomplete, raising hard questions about reliability, equity, and the risks of automating decisions that affect public safety.
What is verified so far
The strongest evidence of a live deployment comes from Southern California. The San Diego County Sheriff’s Department launched an AI call-processing system called Hyper, built specifically to route non-emergency calls. The agency handles up to approximately 400,000 non-emergency calls per year, and Hyper is designed to manage that volume so that dispatchers can prioritize true emergencies. The system’s stated goal is to cut wait times for routine inquiries, such as reports of lost property or noise complaints, that do not require an immediate sworn-officer response.
San Diego’s deployment is significant because it represents an operational, agency-backed rollout rather than a lab experiment. But the department has not yet published post-launch performance metrics, such as average wait-time reductions or error rates in call routing. That gap limits any confident assessment of whether Hyper is meeting its goals, even as it signals that large public-safety agencies are willing to experiment with AI at scale.
On the research side, two academic projects offer technical blueprints for how AI triage could work more broadly. A team working with the Metro Nashville Department of Emergency Communications developed a system called Auto311, detailed in a preprint hosted on arXiv. The researchers built their model on a dataset of 11,796 call recordings and designed it to predict incident types, generate reports, and use confidence-guided dialogue to determine whether a call requires a human dispatcher or can be resolved automatically. The confidence-scoring mechanism is the key innovation: rather than making a binary emergency-or-not decision, the system gauges its own certainty, and escalates ambiguous calls to a live operator.
Separately, a proof-of-concept described in the journal Policing and Society tested natural language processing on 911 and police calls for service in collaboration with Seattle 911. That system ingests calls directly from a 911 phone system and maps call types to response tiers, essentially sorting incoming requests by urgency so dispatchers see the most critical cases first. The Seattle project remains a proof-of-concept, not a permanent deployment, but it demonstrates that NLP-based triage can function within real call-center infrastructure and interface with existing computer-aided dispatch tools.
A related effort also involving Metro Nashville tested a generative-AI-powered training system for 911 calltakers. That system was deployed over six months with 190 users who completed 1,120 training sessions, according to a separate preprint. The training tool simulates caller interactions so new dispatchers can practice before handling real emergencies, exposing them to a wide range of scenarios in a low-risk environment. While this is not a triage system itself, it shows how AI is entering the 911 ecosystem through multiple doors: live call routing, incident classification, and workforce preparation.
What remains uncertain
The most pressing unknown is whether these systems perform equitably across different populations and regions. The Auto311 dataset of 11,796 recordings was drawn from Nashville, and no published analysis addresses how well models trained on that data would handle callers with different regional accents, languages, or speech patterns. AI speech-recognition tools have a documented history of higher error rates for speakers of non-standard dialects, and a 911 triage mistake carries far greater consequences than a misheard voice command on a smartphone.
The generative-AI training deployment in Nashville surfaced practical reliability challenges during its six-month run. The preprint describes difficulties with handling accents and emotionally distressed callers, situations that are routine in real 911 environments. If a training simulator struggles with these variables, a live triage system would face the same problems at higher stakes. No independent audit of any of these systems has been published, and no agency has released data on false-negative rates, meaning calls that should have been flagged as emergencies but were not.
San Diego’s Hyper system presents a different kind of uncertainty. The agency’s announcement frames the tool around non-emergency calls only, which limits the risk profile. A misrouted noise complaint is a far less dangerous error than a misclassified assault report. Still, the boundary between emergency and non-emergency is not always obvious at the moment a call arrives, and the department has not disclosed how Hyper handles edge cases or what fallback protocols exist when the AI’s classification is ambiguous. Without transparency into override rates or supervisor reviews, it is difficult for outside observers to judge whether the safeguards are sufficient.
The Seattle proof-of-concept, published in a peer-reviewed journal, offers the most rigorous methodology of the three projects. But it remains exactly that: a proof-of-concept. No follow-up research or long-term outcome data from Seattle 911 has been published. Without sustained evaluation, it is impossible to know whether the system’s triage accuracy holds up over months of real-world use or whether dispatchers trust it enough to change their behavior. In practice, dispatcher skepticism can blunt the impact of even accurate tools if staff treat AI suggestions as optional or distracting.
How to read the evidence
Readers should distinguish between three types of evidence in this space. The San Diego deployment is a primary operational source, an agency putting AI into production and publicly announcing it. The Auto311 and Seattle projects are primary research sources, meaning they describe systems tested under controlled or semi-controlled conditions and documented in academic papers. The Nashville training tool sits between the two. It was deployed in a real workplace but for training, not live triage, so its stakes and evaluation criteria differ.
None of these sources provide the kind of outcome data that would settle the central question of whether AI triage actually reduces emergency response times or improves caller outcomes. San Diego’s 400,000 annual non-emergency calls offer a large enough sample to generate meaningful performance data, but that data has not been released. The Auto311 and training preprints are available through the arXiv member-supported platform, which hosts early-stage research prior to journal peer review. That status matters: preprints can be revised, challenged, or contradicted as more evidence emerges.
For readers trying to assess credibility, it helps to understand how arXiv itself operates. The repository, which is sustained in part by donor contributions, provides rapid dissemination of technical work but does not replace the scrutiny of formal peer review. Its editorial checks focus on basic relevance and format rather than on validating results. As arXiv explains in its own help guidance, responsibility for the accuracy and interpretation of posted manuscripts rests primarily with the authors and the broader research community that responds to them.
Institutional context also matters. The arXiv service is operated with support from Cornell University, a detail that underscores both its academic roots and its distance from the local agencies experimenting with AI in 911 centers. A university-backed repository can elevate visibility for technical work like Auto311, but it does not guarantee that public-safety departments will adopt those tools, nor that they will share real-world performance data once they do.
What to watch next
For now, the evidence base around AI in 911 call centers is thin but rapidly evolving. The San Diego Sheriff’s Department has taken the most concrete step by moving Hyper into day-to-day use for non-emergency calls, yet it has not released the metrics that would show whether callers are actually getting faster service. Academic teams in Nashville and Seattle have demonstrated that machine learning models can classify calls, generate summaries, and prioritize queues, but they have not yet shown that these capabilities translate into better outcomes for people in crisis.
Key questions remain unanswered. Agencies have not disclosed how they will monitor bias across different communities, what thresholds they use for escalating calls from AI to humans, or how they will communicate these changes to the public. Labor implications are also unclear: AI triage could reduce burnout by filtering routine calls, or it could increase pressure on human dispatchers by concentrating only the most traumatic incidents in their queues.
Future reporting will hinge on whether agencies and researchers publish detailed evaluations, including error rates, demographic breakdowns, and before-and-after comparisons of response times. Until then, the available evidence supports a cautious reading: AI systems are beginning to shape how 911 calls are routed and how dispatchers are trained, but the technology’s real impact on safety, equity, and trust remains an open question that only transparent data and long-term study can answer.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
