{"resourceId":"dhs-oig-26-05-wildfire-sensors-20260427","versions":[{"version":"external-828409e150075cfaa38ca4d17f6ac6c1e3f128aaa5f7e39a7d4fc627c2bafea1","resource":{"id":"dhs-oig-26-05-wildfire-sensors-20260427","title":"DHS OIG: AI wildfire sensors missed fires, posted high false positives, and often lagged 911 alerts","organization":"U.S. Department of Homeland Security Office of Inspector General","sector":"Federal oversight / wildfire detection R&D","geography":"United States (Beta deployments in California, Colorado, Hawaii, Tennessee, and Canada)","publishedAt":"2026-04-27","publicationDate":"2026-04-27","eventDate":null,"sourceName":"DHS Office of Inspector General","sourceLabel":"Federal inspector general final audit report OIG-26-05 on DHS/S&T N5 wildfire sensors","sourceUrl":"https://www.oig.dhs.gov/sites/default/files/assets/2026-04/OIG-26-05-Apr26.pdf","evidenceClass":"government-audit","outcomeClass":"cautionary","topics":["infrastructure","governance-procurement","operating-model","data-security"],"finding":"OIG-26-05 (Apr 27, 2026) finds that DHS S&T’s 2020–2024 N5 Sensors wildfire detection program—using AI/ML on particulate and environmental signals to email/text partners—did not consistently detect fires or provide early warning. At one Beta site (May 2024–Apr 2025), 9 of 13 alerts were false positives; of four fire alerts, three arrived after 911, including a sensor 500 feet from a structure fire that alerted 45 minutes after dispatch. Wind direction/speed blocked smoke from reaching sensors; OIG states even improved AI remains wind-dependent. N5 contract ended Dec 31, 2024; as of Apr 2025 S&T had no plans to further fund the technology. Report contains no recommendations; DHS declined formal comments.","sledRelevance":"Direct federal audit evidence for state/local fire and emergency-management buyers evaluating AI early-warning sensors amid ongoing camera-network procurements. Why now: primary OIG PDF (not the shorter oversight.gov summary already archived) supplies quantified FP/late-alert failure modes that contradict standalone-sensor marketing as ALERTCalifornia/Pano-style camera expansions continue.","evidence":"Opened OIG-26-05 PDF via curl. Highlights, background on ~300 Beta sensors, Beta-site 13-alert breakdown, 500-foot/45-minute example, wind dependence, contract end, and no-recommendations statement taken from the report text.","architectureImplications":"Treat AI sensors as cueing layers fused with cameras/satellite/911, with synchronized alert clocks and mandatory human adjudication before resource orders.","governanceImplications":"Require published FP/miss and lead-time metrics in contracts; do not accept R&D pilots without exit criteria when environmental physics dominate model scores.","securityPrivacyImplications":"Partner alert lists and sensor geolocations are operationally sensitive; audit trails must support after-action without exposing tactics publicly.","caveats":"Single-program sensor tech (not camera AI); Beta partner sample; no OIG recommendations; international Canada sites included; not a PSAP clinical study.","streamIds":["emergency-services"],"roles":{"sales":"Frame the DHS OIG wildfire-sensor audit as cautionary procurement evidence: AI-assisted early-warning sensors can miss fires, generate false positives, and lag 911—especially when wind-dependent—so do not sell standalone sensor AI as a primary detection layer. Discovery should map existing 911/camera workflows, partner alerting SLAs, and whether vendors claim wind-independent performance. Bounded engagement: independent field trial design with false-positive/miss metrics against 911 timestamps. Unsupported claims include guaranteed earlier-than-911 detection or replication of Alpha/Beta marketing results after contract end.","engineering":"Fit is an environmental sensor mesh with AI anomaly scoring plus text/email partner alerts—not a camera network or CAD replacement. Architecture must log alert time vs 911/dispatch, retain false-positive adjudication, and support multi-sensor fusion the OIG notes was envisioned but not fielded in testing. Prerequisites: partner MOUs, wind/meteorology context, and labeled operational data. Constraints: smoke must reach sensors; AI training data hunger; program defunded after Dec 2024. Proof of value is early true-positive rate before 911 and FP rate under partner adjudication—not model accuracy on lab sets alone.","delivery":"Delivery work includes partner alert playbooks, human verification before public warning, and after-action reconciliation of each alert against fire ground truth. Dependencies span S&T/FEMA or successor sponsors, local fire/EM partners, and sensor vendors. Ownership stays with the fire/EM operations center for response decisions. Acceptance criteria: documented FP/FN rates, time-to-alert vs 911, and explicit non-standalone design. Risks include alert fatigue from FPs and false confidence after late detections near structures."},"roleDetails":{"sales":{"customerProblem":"Agencies funding AI wildfire sensors need evidence they actually beat 911 and avoid false alarms under real wind and terrain.","stakeholders":["State/local fire EM partners","USFA/FEMA wildfire programs","S&T or successor R&D sponsors","PSAP/dispatch leadership"],"discoveryQuestions":["What is the measured time from ignition to sensor alert vs first 911?","Who adjudicates false positives before firefighter dispatch?","Is the sensor designed as standalone or fused with cameras/satellite?","What happens when wind blows smoke away from the mesh?"],"valueHypothesis":"AI particulate/gas sensors may add a secondary cue only if field FP/FN and lead-time metrics beat existing 911/camera baselines under local meteorology.","potentialEngagement":"Independent pilot evaluation protocol and after-action metrics workshop—not a sole-source sensor buy based on Alpha/Beta claims.","unsupportedClaims":["Consistent earlier-than-911 detection","AI will self-improve enough to overcome wind dependence","Standalone mesh replaces lookouts or camera networks"]},"engineering":{"fit":"Partner-alerting AI sensor mesh for wildfire anomaly cues with mandatory human verification—not autonomous dispatch.","architectureIntegration":"Sensor telemetry → AI anomaly scoring → text/email alerts → partner console with FP adjudication and time sync to CAD/911.","prerequisites":["Partner MOUs","Labeled operational fire events","Meteorology context","Alert audit store"],"constraints":["Wind-carried smoke dependence","Training-data volume unclear","Contract ended Dec 2024 with no further S&T funding as of Apr 2025"],"security":"Protect partner contact channels and location of sensors; avoid exposing deployment maps that enable tampering.","proofOfValue":["Share of fires alerted before 911","False-positive rate per partner site","Alert latency after known ignition","Missed fires within sensor range"]},"delivery":{"work":"Stand up partner verification SOPs, reconcile each alert to ground truth, and brief leadership on FP/miss patterns before expansion.","dependencies":["Local fire/EM partners","911 timestamp access","Vendor model update process","OIG/audit findings incorporation"],"ownership":"Fire/EM operations owns response decisions; program office owns sensor fleet performance reporting.","skillsAdoption":"Train partners to treat AI alerts as cues requiring confirmation, not automatic resource orders.","governanceCheckpoints":["Pilot go/no-go on FP rate","Monthly alert adjudication review","Public transparency on limitations","Defund/exit criteria"],"acceptanceCriteria":["Documented FP and miss rates","Time-to-alert vs 911 for each true fire","Written non-standalone architecture"],"risks":["Alert fatigue","Late alerts after structure involvement","Overfitting claims without wind controls"]}},"retrievedAt":"2026-09-25T22:04:33-05:00","enrichedAt":"2026-09-25T22:04:33-05:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"False positives and late alerts increase partner workload and can erode trust among already-thin wildfire duty officers.","procurementImplications":"Prefer competitively evaluated, multi-sensor architectures with measurable early-warning SLAs; avoid sole-source renewals after failed Beta consistency.","operatingModelImplications":"Keep 911 and human lookouts/camera watch as primary detection paths; AI sensor alerts remain secondary cues under SOPs.","sourceVerification":{"openedUrl":"https://www.oig.dhs.gov/sites/default/files/assets/2026-04/OIG-26-05-Apr26.pdf","referenceExcerpt":"Of the 13 alerts, 9 were false positives; the local partner did not identify any fires.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}