Public Safety & Emergency ServicesIssue 555 records
Emergency Services
The Seattle Times takes a detailed look at an AI nonemergency line. At Snohomish County 911, Aurelian's Ava cut hold times from more than 10 minutes to 47 seconds and passes about 41% of calls to dispatchers. However, missed-emergency and error figures come only from the vendor, the agency carries liability and the deal was sole-source. A Korean simulation found a protocol-grounded ChatGPT matched 119 paramedics' triage accuracy. Seoul's Yongsan-gu runs tiered AI crowd-density alerts at Itaewon. A rural Korean county credits AI camera prioritisation with early fire detection, and Israel's AI wildfire drone completed a first flight test.
- Evidence records
- 5
- Cross-source patterns
- 2
- Evidence classes
- 3 independent reporting1 academic research1 vendor claim
- Outcomes
- 3 emerging1 mixed1 effective
- Source freshness
- 5 new this fortnight
- Research completed
- 2026-11-02
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
Emergency AI is being deployed as a triage layer whose escalation threshold decides what humans see
SNO911 tunes Ava's 'resistance' so that about 41% of calls go to nonemergency dispatchers and nearly 10% to the emergency queue. Yongsan-gu's four-tier density scale pushes footage and alerts to its disaster centre, police and fire once the danger level is reached. Changnyeong's system automatically puts AI-detected events on operators' screens in place of sequential viewing. In each case the threshold is the main operational control, yet none of the sources reports how often items below the threshold turned out to matter.
Operating questionWho sets the threshold that decides which calls, crowds or camera events reach a human, and how are changes logged and reviewed?
Supporting evidenceSnohomish County 911 (SNO911); Aurelian (vendor of Ava and Cora); Washington State 911 Advisory Committee RAIN subcommitteeYongsan-gu District Office (Seoul); Ministry of the Interior and Safety mass-gathering notification systemChangnyeong County Video Information Integrated Control Center (South Gyeongsang Province, South Korea)
Accuracy evidence for emergency AI is still simulated, vendor-reported or absent
Ava's missed-emergency count and 5–8% word error rate come only from Aurelian, and no competing tool's error rates are public. The Korean triage study's parity with paramedics comes from 100 simulated scenarios. The Technion drone system has completed one flight test with no published detection metrics. Each source signals promise, but none provides an independent field measure of misses, so buyers would need to build that measurement in themselves.
Operating questionBefore relying on an AI tool, what independent field measure of misses and false alarms will you require, and who will collect it?
Supporting evidenceSnohomish County 911 (SNO911); Aurelian (vendor of Ava and Cora); Washington State 911 Advisory Committee RAIN subcommitteeKi-Ung Kan (Cheongju Dongbu Fire Station, Chungcheongbuk-do Fire Service HQ) and Byungwoo Kim (Korea National University of Transportation); Korean Institute of Fire Science and EngineeringTechnion – Israel Institute of Technology; Israel Aerospace Industries; Israel Fire and Rescue Authority; Ministry of National Security; Israel Innovation Authority
Evidence ledger
Full record · every source keeps its link and limitations
Seattle Times: Snohomish County's AI nonemergency agent cut hold times from 10+ minutes to 47 seconds, but missed-emergency counts come only from the vendor and the agency carries liability
The Seattle Times (Oct. 31) reports that SNO911 has used Aurelian's AI agent Ava on its nonemergency line since November 2024 while 911 calls still go to people; the reporter says the technology is used in 20 or more states and at least eight Washington counties. About 41% of SNO911 calls that begin with Ava are transferred to nonemergency dispatchers and nearly 10% go to the emergency queue; agencies tune a 'resistance' setting that controls transfers. The agency says nonemergency hold times that regularly exceeded 10 minutes fell to an average of 47 seconds over the past year. SNO911 bought Ava without competitive bids and was billed $118,944 for 2026 (about 192,000 projected calls at 61 cents each), with a discount for dispatcher feedback. Under the contract SNO911 bears liability for dispatcher mistakes based on Aurelian recommendations, including the Cora emergency-call co-pilot adopted in January. Aurelian's CEO says Ava mistook an emergency for a nonemergency once (2024) and reports a 5% to 8% word error rate; the article notes those are company-only figures, that no error rates for competing tools (GovWorx NEAL, Versaterm CallTriage, Axon Prepared AI) are public, and that no state or federal law requires reporting missed emergencies. A Washington 911 Advisory Committee subcommittee finished drafting voluntary 'responsible AI in 911' practices in October.
Why it matters, evidence and limitations
- Why it matters
- It's an unusually detailed public account of one US PSAP's AI nonemergency deployment, covering transfer rates, hold-time change, per-call price, sole-source buying, liability terms and the absence of independent accuracy data. State 911 offices and county boards can use it to set contract and disclosure terms.
- Evidence and measured results
- Independent newspaper investigation that draws on interviews, an invoice and the contract. Hold-time and transfer figures are agency-reported, while missed-emergency and word-error figures come from the vendor. The article reports no independent audit, sample of call outcomes or demographic accuracy data. It cites a 2022 University of Washington study on speech-recognition disparities as context, and that study isn't specific to Ava.
- Limitations and uncertainty
- Single-agency focus. Benefits are agency-reported and error figures are vendor-reported. Aurelian didn't clarify how many Washington counties use Ava. A 47-second average hold time isn't the same as caller outcomes. Interpretation: results may not transfer to centres with different call mixes or staffing.
Fire Science and Engineering: ChatGPT given Pre-KTAS training materials matched 119 paramedics' triage accuracy (0.81 vs 0.80) on 100 simulated scenarios, with lower specificity
A simulation study published online Oct. 31 compared a non-trained LLM, an LLM given Pre-KTAS (Prehospital Korean Triage and Acuity Scale) educational materials, and 10 active-duty 119 paramedics on 100 scenarios. Each LLM ran every scenario 30 times. Overall accuracy was 0.69 ± 0.10 for the non-trained LLM, 0.81 ± 0.03 for the trained LLM and 0.80 ± 0.03 for paramedics. The trained LLM and the paramedics were more accurate than the non-trained model, with no significant difference between them. In the Korean abstract, the trained LLM had higher consistency (ICC 0.940 vs 0.836) and higher agreement with paramedics (kappa 0.66 vs 0.56), along with the lowest under-triage rate (6.1%) and the highest sensitivity (0.95) and NPV (0.97). Paramedics had the highest specificity (0.96) and PPV (0.93). AUC was 0.934 for both the trained LLM and the paramedics. The authors say field validation and institutional and ethical guidance are needed before clinical use.
Why it matters, evidence and limitations
- Why it matters
- It's peer-reviewed, practitioner-led evidence that a general-purpose LLM given protocol materials can match paramedics on a standard triage scale in simulation. US EMS agencies looking at LLM triage aids can use it to size a validation study, since the result is about protocol grounding and not about field performance.
- Evidence and measured results
- Simulation-based comparative study (100 written scenarios, 30 repetitions per LLM, 10 paramedics) analysed with Welch's ANOVA, Games-Howell, ICC, kappa, chi-square and ROC. I inspected only the article landing page with its English and Korean abstracts. I didn't read the full-text PDF, so the scenario design, ChatGPT version and prompts weren't checked.
- Limitations and uncertainty
- Simulated text scenarios, not live calls or field encounters. Small paramedic comparison group (10). Korean triage scale. Only the abstract was inspected. Interpretation: transferability to US protocols (for example ESI or MPDS) is untested.
Seoul's Yongsan-gu runs 218 AI CCTV cameras at 59 Itaewon sites with a four-tier crowd-density scale that alerts its disaster safety centre, police and fire
Ahead of the fourth anniversary of the Itaewon crowd crush, Yongsan-gu operates 218 AI smart CCTV cameras at 59 locations in and around Itaewon. They count people and estimate crowd density per unit area in real time, and the data is classified on a four-tier congestion standard. When the danger threshold is reached, footage is prioritised on the control-room display and alerts go to the district disaster safety operations centre and to police and fire units. Three electronic boards show crowd information and evacuation directions in Korean and English. The district combines the video analytics with pedestrian-flow data collected every five minutes from mobile carrier base stations, subway exit counts and Seoul's urban data platform. When a crowd-risk alert arrives from the national mass-gathering notification system, operators verify it through the AI CCTV footage. Monitoring is intensified for Oct. 23 to Nov. 1, and the district plans to run the system year-round for large gatherings.
Why it matters, evidence and limitations
- Why it matters
- US emergency managers planning large events (parades, festivals, stadium egress) can compare a concrete multi-source crowd-density warning design, with thresholds, alert routing to fire and police, and public displays, against their own crowd-safety plans.
- Evidence and measured results
- News report, AI-translated, based on district statements and a mayoral site visit. It gives no accuracy, false-alarm or response-time data and doesn't report any intervention triggered by the system.
- Limitations and uncertainty
- Planned and operating configuration only, with no measured outcomes. AI-translated article. Interpretation: South Korean legal and telecom data-sharing conditions differ from US jurisdictions.
Changnyeong County, Korea: AI auto-prioritised CCTV events replace sequential monitoring of 1,515 cameras; county reports two early fire detections with engines on scene within 10 minutes
Changnyeong County's control centre has 16 operators in four teams on three shifts monitoring about 1,515 CCTV cameras, more than 300 per operator. The intelligent CCTV system introduced in 2019 had low accuracy and frequent false alarms from animals and weather. The county built a new monitoring management system that automatically puts AI-detected unusual events on operators' screens, which it calls a national first, and says this raised monitoring efficiency more than tenfold over the previous approach. It reports that in March and July smart monitoring detected township fires before they spread and fire engines reached the scenes within 10 minutes. The system also flagged people at risk of suicide near a bridge and reservoirs, and police reached them within five minutes. The county says initial responses to major incidents more than doubled year on year. It plans an AI-based real-time missing-person tracking system.
Why it matters, evidence and limitations
- Why it matters
- A small rural county reports AI-prioritised video monitoring catching fires early, and it names the false-alarm failure of an earlier system. That makes it a useful case for US counties weighing AI camera analytics for wildfire or structure-fire detection alongside existing staffing. It's cross-tagged public-safety because the same system supports police rescues and suspect tracking.
- Evidence and measured results
- County-reported outcomes via news report. 'More than tenfold' efficiency and 'more than doubled' initial responses have no stated metric definitions or baselines. Fire cases are described anecdotally (two months), with no false-alarm rate for the new system.
- Limitations and uncertainty
- Self-reported, award-oriented county account (Korea Knowledge Awards excellence award) with undefined metrics. Interpretation: an archived Korea Forest Service dataset showed high error-detection shares for AI wildfire CCTV in the same province, so precision claims need verification.
Technion, IAI and Israel Fire and Rescue Authority complete first flight test of an AI drone system to detect wildfires within minutes of ignition
Technion, the Ministry of National Security, the Israel Innovation Authority, IAI and the Israel Fire and Rescue Authority are developing a drone-based system to detect fires within minutes of ignition, and it has completed its first successful flight test. The system runs on IAI's APUS 25 VTOL drone with day and night cameras and infrared sensors. Technion's AI and image-processing algorithms analyse the data in real time, identify and localise fire outbreaks and send alerts to the Fire and Rescue Authority command centre. The announcement says the drone can cover hundreds of square kilometres per flight and stay airborne for several hours.
Why it matters, evidence and limitations
- Why it matters
- It's a government-backed program pairing a national fire service with university AI for airborne early detection. US state forestry and fire agencies comparing camera, satellite and drone detection can use it as a reference point, but it isn't performance evidence. It's dated Oct. 28 and was first surfaced in this run's discovery.
- Evidence and measured results
- Developer and partner announcement republished by a community news site, describing one flight test. It gives no detection accuracy, false-alarm rate, time-to-detection or test conditions.
- Limitations and uncertainty
- Promotional, single milestone with no metrics. Interpretation: military-grade drone platforms and Israeli airspace rules may not transfer to US agencies.
How to read this edition
Source findings, measured results and limitations come from the cited publications. Patterns, operating questions, role takeaways and implementation considerations are Lighthouse Advisory interpretation, stated as questions to validate locally rather than guaranteed outcomes. Vendor and operator claims are labeled as claims. Full research method.
- Independent reporting
- Independent reporting with attributable sources but without a formal evaluation design.
- Academic research
- Research produced through an academic institution or peer-reviewed venue.
- Vendor claim
- A supplier-provided assertion that has not been upgraded to independent evidence.