{"resourceId":"vago-ai-distracted-driver-camera-audit","versions":[{"version":"external-4bca3c55fbdede2b9cdb3e0443e0fd4e8799b6fa09800a3fbcd3104e3ee6c8cd","resource":{"id":"vago-ai-distracted-driver-camera-audit","title":"Reducing the Harm Caused by Distracted Drivers (Victorian Auditor-General audit of an AI camera program)","organization":"Victorian Auditor-General's Office (VAGO)","sector":"State audit office (international comparator)","geography":"Victoria, Australia","publishedAt":"Tabled 10 June 2026","publicationDate":"2026-06-10","eventDate":null,"sourceName":"Victorian Auditor-General's Office","sourceLabel":"VAGO audit report","sourceUrl":"https://www.audit.vic.gov.au/report/reducing-harm-caused-distracted-drivers","evidenceClass":"government-audit","outcomeClass":"cautionary","topics":["governance-procurement","data-security","operating-model"],"finding":"VAGO found the department does not know whether the AI distracted-driver and seatbelt camera program reduces crashes, serious injuries or deaths: no baseline data, changed benefit measures and inconsistent reporting. Image verification works as intended (per 1,000 vehicles photographed, AI passed 147 to human review and 6 infringements resulted), but images were kept longer than regulations allow, the vendor archived images for 7 years until February 2026, self-reported vendor KPIs were not verified, and the department could not confirm which AI software version was deployed from March 2024 until December 2025.","sledRelevance":"Why now: U.S. agencies are launching or expanding AI camera enforcement this month (MBTA education period from October 1; Chicago citywide expansion October 14). This audit is a detailed public-sector checklist for evaluation baselines, retention, vendor KPI verification and model version control that U.S. programs can apply before scaling.","evidence":"Independent government audit with program data (308 sites, 188,493 infringements, July 2023 to September 2025) and five recommendations (four accepted in full, one in part).","architectureImplications":"AI filtering followed by multi-stage human review; a vendor cloud archive caused a retention breach, so backup and archive design must honor deletion rules.","governanceImplications":"Evaluation must be designed before launch with baselines; model version changes need records; vendor self-reported KPIs need independent verification.","securityPrivacyImplications":"Retention breaches occurred in staff email, internal files and vendor backups; deletion must cover every copy.","caveats":"Australian legal context; the audit concerns evaluation and administration, not detection accuracy itself; published June 2026, before this edition.","streamIds":["transportation"],"roles":{"sales":"This audit gives buyers of AI camera enforcement a ready-made list of what goes wrong when evaluation and retention are afterthoughts. For U.S. agencies launching programs now, the bounded offer is pre-launch evaluation design and contract controls. Do not cite it as evidence that AI cameras fail at detection; VAGO found verification worked. The finding is that the agency could not show road-safety benefit.","engineering":"Engineering takeaways: deletion rules must reach every copy, including vendor cloud backups and staff email; model version deployment must be logged; KPI data should be independently verifiable. AI filtering cut review to 147 of every 1,000 images, which sizes downstream review. Build evidence pipelines with retention enforcement and audit trails from day one rather than retrofitting after an audit.","delivery":"Delivery should establish baselines and a benefits plan before go-live, not two years in. Set a fixed cadence for governance committee reporting, verify vendor KPIs by sampling, and record every model change. These proposals draw on VAGO's recommendations, not observed U.S. outcomes. Assign one owner for retention compliance across department and vendor systems."},"roleDetails":{"sales":{"customerProblem":"Agencies deploying AI enforcement cannot later prove safety benefit or privacy compliance without upfront design.","stakeholders":["Transportation safety program leads","Privacy officers","Contract managers","Auditors"],"discoveryQuestions":["What baseline crash and behavior data exists before launch?","Can you independently verify vendor KPIs?","How are AI software versions approved and recorded?"],"valueHypothesis":"Pre-launch evaluation and contract controls could prevent audit findings like VAGO's.","potentialEngagement":"A bounded pre-launch evaluation and contract-controls review for an AI camera program.","unsupportedClaims":["That AI cameras reduce crashes","That AI detection is inaccurate","Any U.S. revenue projection"]},"engineering":{"fit":"Fits retention enforcement, audit logging and evaluation data pipelines.","architectureIntegration":"AI image filtering, multi-stage human review, police adjudication and a vendor cloud archive.","prerequisites":["Retention policy mapped to every storage location","Model version registry","Agency access to KPI data"],"constraints":["Jurisdiction-specific retention regulations","Vendor-operated AI and systems"],"security":"Ensure deletion propagates to backups and archives and restrict image sharing by email.","proofOfValue":["Retention compliance scan across systems","Independent sample check of vendor KPIs","Version log reconciliation"]},"delivery":{"work":"Build evaluation baselines, retention compliance and vendor verification into program launch.","dependencies":["Crash data attribution method","Vendor cooperation","Governance committee"],"ownership":"Program-owning department with vendor contract management.","skillsAdoption":"Staff training on retention rules and evaluation methods.","governanceCheckpoints":["Benefits plan approved before launch","Quarterly committee reporting","Change approval for AI versions"],"acceptanceCriteria":["Proposed: baseline dataset documented before go-live","Proposed: no images retained beyond regulatory limits in an audit sample"],"risks":["Unmeasurable benefits","Privacy breaches via backups","Unverified vendor performance"]}},"retrievedAt":"2026-10-29T22:05:00-05:00","enrichedAt":"2026-10-29T22:12:23-05:00","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Multi-stage human verification is labor-intensive; AI upgrades aim to reduce that load.","procurementImplications":"Contracts should include independent KPI verification rights, archive and retention obligations, and model-version change control.","operatingModelImplications":"Independent third-party testing of AI changes was noted as good practice; an evaluation commissioned two years late was not.","sourceVerification":{"openedUrl":"https://www.audit.vic.gov.au/report/reducing-harm-caused-distracted-drivers","referenceExcerpt":"does not know if or how the program is reducing road accidents, serious injuries and deaths","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}