Public Safety & Emergency ServicesIssue 155 records
Emergency Services
Five newly catalogued sources after the 2026-09-19 edition: a Sept 7 Swansea/EMRTS ethnographic preprint placing algorithmic assist at the early 'decision-to-view' in critical-care dispatch rather than final send; Aug 12 Gambit interview bounding New Orleans 911 AI to near-crash duplicates without dispatch or priority powers; Aug 10 explainx scrutiny that the oft-cited zero false-positive/negative claim still lacks an independent audit; Aug 30 TENS analysis that ALERTCalifornia wildfire AI does not declare emergencies or dispatch and needs three-clock KPI hygiene; and the Aug 4 AIDE Way Forward strategy report documenting a 1,179-product landscape alongside a capacity-and-trust adoption gap with six recommended actions. Two patterns require independent 911 AI error audits before scope expansion and prefer early-assist/HITL verification over autonomous send or fire declaration. No new controlled U.S. clinical EMS outcome trial dated after 2026-09-19 is claimed.
- Evidence records
- 5
- Cross-source patterns
- 2
- Evidence classes
- 3 independent reporting1 academic research1 independent research
- Outcomes
- 4 emerging1 cautionary
- Source freshness
- 4 recent1 new this fortnight
- Research completed
- 2026-09-21
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
Narrow 911 AI still needs published independent error audits before scope expansion
Gambit's Fasold interview bounds New Orleans AI to near-crash duplicates without dispatch/priority powers, while explainx flags that the zero false-positive/negative claim lacks an independent third-party audit and that live use ran for years with limited public disclosure. Together they counsel SLED PSAPs to freeze expansion until sampled FP/FN results and caller disclosure are public.
Operating questionBefore expanding beyond a single 911 AI use case, what independent sample and disclosure artifacts must the PSAP publish?
Supporting evidenceOrleans Parish Communications District (OPCD)explainx.ai analysis of OPCD / Carbyne (Axon) deployment
Prefer early-workflow assist and human verification over autonomous send/declare
The EMRTS ethnography places algorithmic help at the early decision-to-view rather than final dispatch, and TENS stresses ALERTCalifornia does not declare emergencies or dispatch crews—watchstanders verify. SLED buyers should scope AI as sift/verify layers with human authority on launch and public warning.
Operating questionWhich workflow step is AI allowed to assist, and which life-safety actions remain exclusively human?
Supporting evidenceSwansea University / EMRTS (Swansea Bay UHB)ALERTCalifornia / CAL FIRE (as analyzed by TENS Magazine)
Evidence ledger
Full record · every source keeps its link and limitations
EMRTS ethnographic study: place algorithmic assist at early 'decision-to-view' in critical-care dispatch
Ethnographic immersion and 22 operational hours of structured observation across three day and two night shifts at a Welsh EMRTS pre-hospital critical-care dispatch hub map a multi-step workflow and conclude that the early 'decision to view' a call from the dynamic list is the most promising place for unobtrusive algorithmic assistance—an unobtrusive score inside the existing call list—while later dispatch decisions depend on crew skill knowledge and interactive clinical judgment that staff judged poorly suited to algorithms. Why now: after the 2026-09-19 LogiDebrief post-call QA package, this Sept 7 paper is the first newly catalogued fieldwork that tells SLED EMS/ECC buyers where NOT to put live AI in critical-care dispatch, distinct from archived post-call QA metrics.
Why it matters, evidence and limitations
- Why it matters
- Transferable design-placement evidence for U.S. SLED EMS and ECC AI copilots: prefer early list-triage assist over autonomous send/no-send decisions, with clear UK-to-U.S. protocol and governance transferability limits.
- Evidence and measured results
- Opened arXiv HTML and PDF (2609.07705v1, submitted 7 Sep 2026). Methods cite immersion plus 22 operational hours over three day and two night shifts. Conclusions state the technological intervention focus will be the early decision to view, implemented as an unobtrusive simple score in the existing call list. Staff skepticism about algorithmic situational awareness for later dispatch steps documented in section 3.3.3–3.3.4.
- Limitations and uncertainty
- UK EMRTS critical-care hub ethnography, not a U.S. PSAP RCT; no deployed accuracy metrics for the proposed score; transferability limited by protocol and staffing differences.
Gambit: New Orleans 911 AI limited to near-incident crash callbacks; cannot dispatch or prioritize
Gambit's Aug 12 2026 interview with OPCD Executive Director Karl Fasold clarifies that New Orleans' 911 AI (in use since 2023 for a narrow use case) answers only callers within about 200 meters of a known car accident, is not a large language model like ChatGPT, and cannot end calls, dispatch responders, or prioritize calls—while a limited 311 AI began more recently. Why now: the July fox8live item is already archived; this local primary interview supplies the operational boundary language SLED PSAPs need when vendors market 'AI answers 911,' and pairs with post–2026-09-19 scrutiny of unverified accuracy claims.
Why it matters, evidence and limitations
- Why it matters
- U.S. local PSAP deployment of narrowly scoped duplicate-incident AI with explicit non-dispatch limits—directly relevant to SLED 911 staffing-gap procurements.
- Evidence and measured results
- Opened Gambit HTML dated Aug 12, 2026 (editor's note correcting 911 AI start to 2023 vs 311 AI). Fasold quotes and statements that the system cannot end calls, dispatch, or prioritize; ~200-meter proximity trigger; not an LLM; dialect/street-name training cited in surrounding coverage context.
- Limitations and uncertainty
- Single local interview piece; performance metrics in this article are qualitative; does not independently audit error rates.
explainx: New Orleans Carbyne 911 AI narrow but unaudited—zero FP/FN claim remains unverified
An Aug 10 2026 explainx deep dive argues New Orleans has run Carbyne AI on live 911 traffic since 2023 for crash-call deduplication (with ~200 m matching and human escalation), confirmed publicly around Aug 6 2026 after limited prior disclosure, and itemizes what remains unconfirmed—including any independent third-party audit of the oft-repeated zero false-positive / zero false-negative claim from the initial fully monitored window. Why now: pairs with Gambit's scope interview to give SLED buyers a scrutiny checklist after non-emergency AVA and CommsCoach themes already in the archive.
Why it matters, evidence and limitations
- Why it matters
- Independent scrutiny template for U.S. SLED PSAPs evaluating vendor accuracy claims on live emergency call AI.
- Evidence and measured results
- Opened explainx HTML dated Aug 10, 2026. TL;DR table lists unconfirmed independent audit of zero FP/FN, full incident-level error rate since 2023, and contract terms beyond a reported ~$600K/year figure. Describes classifier-style architecture, 100% human monitoring at launch, and escalation on uncertainty.
- Limitations and uncertainty
- Secondary analysis synthesizing press and official statements; not a formal government audit; some dollar and staffing figures are reported, not independently verified here.
TENS: ALERTCalifornia wildfire AI is fastest when humans verify—system does not self-dispatch
TENS Magazine's Aug 30 2026 analysis of ALERTCalifornia states the AI filters camera streams for possible smoke and brings candidates to trained watchstanders—explicitly noting the system does not declare an emergency or dispatch crews by itself—and introduces a 'three clocks' framing (model, verification, response) while cautioning not to merge Microsoft early-period and ALERTCalifornia first-season detection statistics into one rate. Why now: after the 2026-09-18/19 Teton ALERTWest camera edition, this piece supplies transferable HITL and metrics-hygiene language for SLED wildfire camera programs.
Why it matters, evidence and limitations
- Why it matters
- Operational doctrine for U.S. SLED fire/EM agencies deploying AI wildfire cameras: verification before response and honest separation of detection KPIs.
- Evidence and measured results
- Opened TENS HTML dated Aug 30 2026. States system does not declare emergencies or dispatch; describes watchstander confirmation; notes clouds/fog/dust false cues; cites Microsoft 77-incident early window and ALERTCalifornia >1,200 first-season detections with >30% beating 911—warning not to merge denominators.
- Limitations and uncertainty
- Secondary magazine analysis; performance figures attributed to Microsoft/ALERTCalifornia statements across different periods; not a controlled trial.
AIDE Way Forward: EM AI demand is clear but adoption needs capacity, trust, and six actions
The Aug 4 2026 AIDE Report 'AI for Disasters + Emergencies: A Way Forward' synthesizes Aspen Digital practitioner workshops and RAND's landscape of 1,179 AI-enabled products from 717 companies across 45 EM task areas, arguing useful tools and use cases exist but many agencies lack capacity and trust for responsible adoption—the 'last mile' gap—and recommends six interconnected actions (summit, education, playbook, capacity-building pilots, research, alliance). Why now: the Sept playbook PDF is already archived; this companion strategy document was not, and it reframes SLED EM buying after FEMA NGWS AI eligibility surfaced in the 2026-09-19 edition.
Why it matters, evidence and limitations
- Why it matters
- National independent roadmap for U.S. SLED emergency managers procuring and governing AI without mistaking product volume for operational readiness.
- Evidence and measured results
- Opened AIDE Way Forward PDF dated August 4, 2026. Documents RAND count of 1,179 products / 717 companies / 45 task areas; states adoption-gap language that tools and use cases exist while many agencies lack capacity and trust; lists six next-phase actions.
- Limitations and uncertainty
- Strategy synthesis, not a product evaluation; does not rank tools; companion RAND report explicitly declines product scoring.
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.
- Independent research
- Research conducted outside the implementing organization.