Public Safety & Emergency ServicesIssue 195 records
Emergency Services
With non-emergency AI answering spreading after CSPD’s Prepared launch, today’s edition pairs measured/governance reporting from Richmond and Western Massachusetts with Seattle council oversight of Corti medical-prompt AI, and cautionary multimodal science from SAFIRE (61.9% MLLM fire/smoke VQA) and Nature’s DisasTeller (error-propagating LVLM disaster agents). Net signal: workload tools are real, but disclosure, labeled transcripts, public metrics, and staged HITL remain the operating controls.
- Evidence records
- 5
- Cross-source patterns
- 2
- Evidence classes
- 3 independent reporting2 academic research
- Outcomes
- 3 cautionary1 mixed1 emerging
- Source freshness
- 4 recent1 older, newly relevant
- Research completed
- 2026-09-25
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
Non-emergency AI answering spreads faster than disclosure and shared outcome metrics
Richmond reports partial containment and better 911 answer-within-15s figures without caller AI notice; Westcomm is preparing Prepared/Axon answering with scram controls amid ACLU transcript worries; Seattle’s council hearing shows even HITL medical-prompt AI ran years without public metrics. Agencies need disclosure, labeled transcripts, and published SLAs alongside workload claims.
Operating questionBefore expanding AI on business or diversion lines, can the PSAP show caller notice, escalation proof, and public answer-time or wait metrics?
Supporting evidenceAxios RichmondGovTech / masslive.com (Tribune Content Agency)The Seattle Times
Multimodal fire/disaster models still fail safety-critical reasoning and can cascade errors
SAFIRE finds open MLLMs average only 61.9% on fire/smoke VQA with weak spatial/general reasoning, while DisasTeller shows early LVLM grading mistakes propagating into allocation and reconstruction drafts. Both point to staged human verification—not autonomous scene declaration or EOC command.
Operating questionWhere in the camera-to-alert or image-to-allocation chain is a human required to catch fluent multimodal errors?
Supporting evidenceKhalifa University / University of Western Australia (arXiv preprint; EMNLP 2026 Findings)Nature Communications (University of Sydney et al.)
Evidence ledger
Full record · every source keeps its link and limitations
Seattle council questions Fire Department on undisclosed Corti AI triage and nurse-line waits
After Times reporting, Seattle’s Public Safety Committee pressed fire officials on AMR’s Texas nurse line and Corti AI that has listened to all medical 911 calls since 2023 and prompts transfers. Chief Scoggins said dispatchers retain final authority; callers were not disclosed AI use; UW is evaluating effectiveness without public metrics shared Tuesday. Council focus is restoring tracking/time standards for nurse-ordered ambulances after the Pamela Hogan case.
Why it matters, evidence and limitations
- Why it matters
- Shows how fire/EMS PSAPs can deploy live AI prompts for medical diversion without caller notice or surveillance-ordinance review—directly relevant to U.S. emergency communications governance, contract backstops, and HITL claims.
- Evidence and measured results
- Primary local investigative reporting on Sep 8, 2026 committee hearing; quotes Scoggins, Councilmember Kettle, and references prior Times series and settled Hogan estate suit. Not a controlled outcome study; diversion increase figures remain disputed elsewhere.
- Limitations and uncertainty
- Single-city reporting; UW evaluation details not released; nurse-line wait tracking remains with AMR, so city-level outcome metrics for diverted callers are incomplete.
Richmond’s Amazon Connect AI handles some non-emergency calls; emergency answer-within-15s rose to ~93%
Richmond’s Amazon Connect non-emergency assistant (live since April 2025; >$52k spent) fully handles ~10–15% of non-emergency calls and routes ~5% elsewhere. Officials link it to freeing 911 capacity: January emergency answer-within-15-seconds ~93% vs ~85% prior year. Callers are not told they speak with AI; Axios found inconsistent behavior including a disconnect and Spanish-to-English handoff issues. NENA advises disclosure for trust.
Why it matters, evidence and limitations
- Why it matters
- Provides rare cost and performance numbers for U.S. city non-emergency AI adjacent to 911 centers—useful benchmark for PSAP staffing relief claims and disclosure norms.
- Evidence and measured results
- Reporter obtained contract/cost figures from city documents and spokesperson; performed five test calls. Emergency answer-time improvement is city-attributed correlation, not a controlled causal study.
- Limitations and uncertainty
- July 2026 reporting (pre-dates last run) newly archived here because non-emergency AI answering is accelerating (CSPD Sep 22) and Richmond supplies measured costs/answer-time figures still missing from recent editions. Not a peer-reviewed evaluation.
Westcomm (Western Mass.) prepares Prepared/Axon AI to answer non-emergency lines amid ACLU transcript concerns
Westcomm plans an AI answerer for non-emergency business-line overload (79,500 business vs ~30,300 911 calls YTD to July 1). Existing AI already does live transcription, translation, radio transcription, and QA assist (~$100k, state 911 grant). New answering system (Prepared/Axon) will include scram-to-human, supervisor monitoring, and human handling of medical/alarm calls. ACLU of Massachusetts warns AI transcripts/summaries could introduce errors into criminal-case evidence; Springfield has no near-term AI answering plans.
Why it matters, evidence and limitations
- Why it matters
- Documents a grant-funded regional PSAP path to non-emergency AI answering with explicit escalation controls and civil-liberties scrutiny—useful for multi-town emergency communications districts.
- Evidence and measured results
- On-record interviews with Westcomm leaders and ACLU counsel; volume and cost figures from agency. System not yet live—outcomes prospective.
- Limitations and uncertainty
- Pre-go-live reporting (Aug 2026) included because it is not archived and newly frames non-emergency answering controls after CSPD’s Sep 22 Prepared launch; no measured containment rates yet.
SAFIRE benchmark: open MLLMs average only 61.9% on fire/smoke safety-critical VQA
SAFIRE evaluates multimodal LLMs on fine-grained fire/smoke understanding across 20 scenarios (83K images; 193K MCVQA). Ten open-source MLLMs (8B–38B) average 61.9% accuracy, with weak spatial correlation (~42.9%) and general reasoning (~53.3%). Lightweight domain adaptation of vision encoders with only 7% labeled data lifts fire-scene classification from 20.1% to 64.5%. Authors stress residual annotation noise and research-only release.
Why it matters, evidence and limitations
- Why it matters
- Sets a quantitative caution for fire/EMS agencies buying multimodal ‘scene understanding’ for cameras, drones, or CAD multimedia—off-the-shelf MLLMs are not yet reliable for safety-critical fire/smoke reasoning.
- Evidence and measured results
- Large constructed benchmark with automated+human QA (human audit ~87.1% on 2,800 items). Not an operational PSAP deployment study; generator-family bias analyzed as limited vs human gap (~82.5% human subset).
- Limitations and uncertainty
- Preprint/Findings paper; synthetic annotation pipeline; not validated on ALERTCalifornia-like live camera nets. Why-now: first major MLLM fire-smoke reasoning benchmark not yet in archive as agencies expand AI cameras.
DisasTeller multi-LVLM disaster agents: early grading errors propagate; HITL audit required
DisasTeller coordinates four LVLM agents (GPT-4o vs Gemma3-27B) for post-disaster assessment, alerts, emergency services, and assignment/reconstruction. GPT-4o intermediate grading averaged 7.3/10 vs 4.6/10 for Gemma3; map annotation 6.0 vs 5.0. Authors show early-stage agent errors can propagate to allocations and plans, and argue intermediate artifacts must enable human audit/intervention. Framework is decision support, not replacement of authority.
Why it matters, evidence and limitations
- Why it matters
- Evidence that multi-agent LVLM ‘EOC copilots’ can accelerate drafting yet still need staged HITL because grading mistakes cascade—material for emergency management AI governance as agentic tools are pitched to EOCs.
- Evidence and measured results
- Peer-reviewed open-access study with repeated runs, human+EvaluatorGPT scoring, and explicit limitation analysis. Case studies use public crisis imagery; allocation numbers are illustrative, not optimized OR solutions.
- Limitations and uncertainty
- Published earlier in 2026 but not archived; newly relevant alongside SAFIRE as multimodal/agentic disaster AI moves from demos to procurement. Not a U.S. operational deployment.
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.