Public Safety & Emergency ServicesIssue 575 records
Public Safety
This public-safety edition looks at who measures police AI. Western Australia Police ended a live facial recognition trial reporting 122 arrests and nine false alerts from 1.38 million scans, but critics dispute its error rate. New Zealand Police began AI interview transcription on a vendor accuracy claim of 80%, requiring officers to record verification. London’s Met opened market engagement for AI CCTV analytics led by evidential integrity. New Jersey Assembly bills would have the attorney general study facial recognition and AI police reports. Recidiviz says Michigan’s prison classification tool is points-based, not AI.
- Evidence records
- 5
- Cross-source patterns
- 2
- Evidence classes
- 4 independent reporting1 vendor claim
- Outcomes
- 4 emerging1 mixed
- Source freshness
- 5 new this fortnight
- Research completed
- 2026-11-03
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
The accuracy numbers behind new police and corrections AI come from the operator or vendor, and the metric choice decides the story
Western Australia Police reported nine false alerts out of 1.38 million faces scanned and called it a 0.00065% error rate, while civil libertarians say the right test is how many of the 122 arrests end in convictions. New Zealand Police adopted interview transcription on Axon’s 80% average accuracy claim, with no local test published. Michigan’s classification pilot claims about 30% better violence prediction, and its provider’s Nov. 2 post explains the method but not that figure. In each case no independent test is public, so buyers should set the metric themselves.
Operating questionFor each AI tool under consideration, which accuracy metric will we report, measured on whose data, and who outside the agency checks it?
Supporting evidenceWestern Australia Police ForceNew Zealand PoliceRecidiviz (nonprofit) with Michigan Department of Corrections
Human review of AI drafts is becoming the standard control, but who does the checking and how it is recorded varies
New Zealand Police require investigators to corroborate auto-transcripts and record that verification, while keeping court transcripts human-made. New Jersey’s A-3076 would require officers to review AI-generated reports, and a critic warns review alone may not be enough. Recidiviz says officers remain authors of AI case notes and must edit substantively, backed by evaluation of every output. The control is the same in all three; the record of review and the checks on reviewers differ, and barristers expect part of the burden to move to lawyers.
Operating questionWhen an officer reviews an AI draft, what evidence of that review is kept, and who audits whether reviews are real?
Supporting evidenceNew Zealand PoliceNew Jersey Legislature (Assembly Public Safety and Preparedness Committee)Recidiviz (nonprofit) with Michigan Department of Corrections
Evidence ledger
Full record · every source keeps its link and limitations
Western Australia Police end a five-month live facial recognition van trial: 1.38 million faces scanned, 122 arrests, nine false alerts, and a disputed error rate
The Guardian reported that Western Australia Police ended a five-month live facial recognition trial on Saturday, Oct. 31, 2026. Police said vans and vehicle-mounted cameras scanned 1,380,882 faces across 100 deployments at 43 locations, matching against an alert list from police databases. Police reported 122 arrests, engagement with 146 registered sex offenders, 10 welfare checks, and nine incorrect alerts, which they described as an error rate of 0.00065% of scans. The commissioner said no one wrongly alerted was arrested but each was approached and questioned. Civil Liberties Australia argued the rate should be judged by how many arrests lead to convictions. The state information commissioner said it was not involved in the trial. Vans are withdrawn while police run an internal evaluation of performance, community impact and safeguards.
Why it matters, evidence and limitations
- Why it matters
- A large, fully reported live facial recognition trial with operator-supplied outcome counts, showing which metrics agencies publish and which they leave out.
- Evidence and measured results
- Opened and read the full Guardian article. All counts are police figures as reported; the internal evaluation and any underlying data were not published.
- Limitations and uncertainty
- Australian law and policing differ from U.S. practice; figures are police self-reports; arrest outcomes and demographic error data were not reported; the vendor is not named in the article.
New Zealand Police begin using Axon Auto-Transcribe for interviews, with vendor-stated 80% average accuracy and barristers warning checking will shift to lawyers
RNZ reported that New Zealand Police began using Axon Auto-Transcribe for interview speech-to-text on Nov. 2, 2026. Axon says the tool averages 80% accuracy on public safety data and over 90% on clearer audio; police acknowledged the transcripts are not always accurate and said investigators must corroborate information and record that verification before acting. Court transcripts will still be prepared and verified by people in the police Transcription Service. Prosecutors may give the auto-transcript, with the recording, to defendants. A barrister warned of breaches and shortcuts in an underfunded system; another said checking would shift to prosecutors and defence counsel, and raised lock-in and data-sovereignty concerns about overseas vendors. RNZ learned of the tests through an Official Information Act release.
Why it matters, evidence and limitations
- Why it matters
- Shows a police agency adopting AI transcription with an explicit human-verification rule while the accuracy figure comes from the vendor.
- Evidence and measured results
- Opened and read the full RNZ article. The 80% and 90% figures are Axon claims as reported; no police evaluation was published.
- Limitations and uncertainty
- Vendor accuracy claim not independently tested; New Zealand privacy and evidence law differ from U.S. state law; time savings are expected, not measured.
London’s Metropolitan Police opens market engagement for AI video analytics to cut manual CCTV review, targeting a three-year contract from March 2027
PublicTechnology reported on Nov. 2, 2026 that the Metropolitan Police published a planning notice for market engagement on AI-enabled CCTV video analytics, to reduce manual review time and turn video into usable intelligence. Requirements quoted include evidential integrity, interoperability, scalability, lawful and ethical AI use, video ingestion, evidential workflows and integration with future digital evidence management. Suppliers had until Nov. 3 to register and must respond to a request for information by Nov. 16. The Met has not committed to procurement but expects an initial three-year contract from March 2027 with a possible two-year extension to 2032; no value is estimated. The article cites an estimate of about 940,000 CCTV cameras in London.
Why it matters, evidence and limitations
- Why it matters
- A large force defining AI video analytics requirements around evidential integrity before buying, a template other agencies can compare against.
- Evidence and measured results
- Opened and read the full PublicTechnology article quoting the notice. The notice itself was not opened.
- Limitations and uncertainty
- Planning stage only; no performance data; secondary report of a notice; UK context.
New Jersey Assembly committee advances bills requiring the attorney general to study police facial recognition and AI-written reports and recommend statewide policy
The Jersey Vindicator reported on Nov. 2, 2026 that two New Jersey Assembly bills cleared the Public Safety and Preparedness Committee on Oct. 19. A-759 would require the attorney general to assess available police facial recognition technology and its accuracy, summarize statewide use, and recommend policy. A-3076 would do the same for AI and call for a statewide policy requiring officers to review AI-generated reports for accuracy and completeness. A sponsor said many departments in his district already use AI report writing without statewide guidance. A Montclair State professor said officer review may not be sufficient because of reliance on AI output, and favoured audits, disclosure to defendants and community input. The article contrasts California’s law (disclaimers, first-draft retention, audit trail) and Utah’s 2025 law. Identical Senate bills are expected.
Why it matters, evidence and limitations
- Why it matters
- Shows a state moving to study-first governance of AI police reports and facial recognition, while peer states already mandate disclosure and audit trails.
- Evidence and measured results
- Opened and read the full article. Bill texts were not opened; descriptions are the reporter’s.
- Limitations and uncertainty
- Bills are in committee stage only; outcomes uncertain; bill texts not reviewed directly.
Recidiviz says Michigan’s new prison classification tool is points-based, not AI, and sets out its review and error-ceiling rules for generative AI case notes
Recidiviz, the nonprofit working with the Michigan Department of Corrections on its classification pilot due Dec. 1, wrote on Nov. 2, 2026 that coverage describing the tool as AI was wrong: both the old and new instruments are rules-based points systems staff can read line by line. It says the redesign analyzed 10 years of Michigan classification data and that MDOC staff experts decided which factors to keep. Recidiviz says it does not use AI to decide custody, parole or release, or to score risk, and that AI only drafts records that staff review and submit. For its generative AI case-note tool, it says staff manually review 100% of outputs in early testing, a second model grades every output later, harmful errors have a strict ceiling, and tools are rolled back if the ceiling is crossed.
Why it matters, evidence and limitations
- Why it matters
- A provider correcting AI framing of a corrections risk tool, and publishing its own generative AI control rules, which buyers can turn into contract terms.
- Evidence and measured results
- Opened and read the full post. All statements are the provider’s own; no error rates, ceiling values or validation results are disclosed. MDOC’s Oct. 28 release was archived previously and is not reused.
- Limitations and uncertainty
- Provider self-description; the 30% improvement claimed by MDOC remains unexplained; no independent validation.
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.
- Vendor claim
- A supplier-provided assertion that has not been upgraded to independent evidence.