Lighthouse AdvisorySLED AI Adoption Intelligence

Public Sector & Government · Issue 02 ·

Public Safety

Today's Milwaukee reporting raises questions about disclosure of earlier facial-recognition use after a moratorium. Recent Thames Valley deployment logs illustrate activity reporting and unresolved data-quality limits. Historical evidence newly added to the archive examines police-report quality and judicial early-adopter experience. Four sources, two interpretations; no proven generalized savings, crime reduction or factual-error improvement. Corrections remains a coverage gap; NCSC coverage is limited to its substantive public summary.

Evidence records
4
Cross-source patterns
2
Evidence classes
1 independent reporting1 academic research1 government evaluation1 public-sector association guidance
Outcomes
2 cautionary1 mixed1 emerging
Source freshness
2 older, newly relevant1 new this fortnight1 undated
Research completed
2026-09-08

Choose a role to see its takeaway beside every record in the ledger.

Synthesis · Lighthouse Advisory interpretation

Patterns across the evidence

2 patterns, each supported by at least two sources

  1. Accountability requires traceable artifacts across organizational handoffs

    Milwaukee's contested disclosure history and the report-quality study's distinction between subjective assessment and factual verification support preserving inspectable source and review records. A human approval field alone does not establish what evidence downstream reviewers received.

    Operating questionCan an authorized reviewer reconstruct the AI contribution, source material and human changes after the output leaves its originating team?

    Supporting evidenceMilwaukee Neighborhood News Service; Devin BlakeIan T. Adams and coauthors

  2. Define the outcome before accepting a favorable proxy

    Police-report ratings, facial-recognition activity and judicial interview perceptions measure different things. Each can inform a pilot, but none alone establishes better justice outcomes or transferable financial benefit.

    Operating questionWhich independently checked service outcome and baseline would justify continuing this specific workflow?

    Supporting evidenceIan T. Adams and coauthorsThames Valley PoliceTRI/NCSC AI Policy Consortium for Law & Courts

Full record · every source keeps its link and limitations

Evidence ledger

4 records
  1. Independent reportingCautionaryNew this fortnight

    Milwaukee reporting highlights legacy facial-recognition disclosure gaps

    Reporting says cases involving earlier facial-recognition use continue after MPD's February moratorium; completeness of disclosure remains contested.

    Milwaukee Neighborhood News Service; Devin BlakeMilwaukee County, Wisconsin, United StatesSeptember 7, 2026

    Why it matters, evidence and limitations
    Why it matters
    A same-day U.S. local development: stopping a tool does not settle handling of evidence already produced.
    Evidence and measured results
    The DA describes matches as leads requiring corroboration and says received information goes to defense. ACLU and defender accounts raise missing-report and system-transparency concerns. MPD did not answer disclosure questions.
    Limitations and uncertainty
    No complete case census, independently adjudicated disclosure failure rate or local algorithm accuracy test. February's exact moratorium date is unspecified. Claims are attributed reporting, not judicial findings.
    Independent reporting with attributed prosecutor, defender and advocate accounts Role takeaways & implementation
  2. Academic researchCautionaryNewly relevant · May 2026

    Blinded police-report study distinguishes perceived quality from factual verification

    AI-assisted reports were less readable, while the primary overall perceived-quality difference was not statistically significant.

    Ian T. Adams and coauthorsUnited States; one police agencyMay 8, 2026

    Why it matters, evidence and limitations
    Why it matters
    New-to-archive quality evidence complements yesterday's Manchester timing trial using related underlying reports; it is not an independent deployment or a new September experiment.
    Evidence and measured results
    The sample contained 20 assisted and 60 conventional reports; 92 raters supplied 354 evaluations covering 79 reports. Flesch scores were 52.28 versus 57.92. Overall quality p=.094; the accuracy-rating subscale p=.038. These are perceptions, not verified factual-error rates.
    Limitations and uncertainty
    Preprint, single agency/tool, small assisted sample, subjective ratings and generic readability metrics. Multiple subscales warrant caution. Methods and discussion differ on when raters were primed about AI; omit strong detection claims.
    Academic preprint; blinded expert evaluation Role takeaways & implementation
  3. Government evaluationMixedUndated source

    Thames Valley register exposes the gap between facial-recognition activity and impact

    Operator logs report deployment activity with variable alerts and disposals, without a counterfactual crime-reduction evaluation.

    Thames Valley PoliceEngland, United Kingdom; limited U.S. transferabilityUndated live HTML; latest listed deployment September 4, 2026

    Why it matters, evidence and limitations
    Why it matters
    New-to-archive recent operational data illustrates what an agency can disclose and what local independent evaluation must still establish.
    Evidence and measured results
    September 4 Wycombe HTML records 33,792 faces seen, eight alerts, zero arrests and six disposals at threshold 0.64. The linked PDF is marked updated September 1 and excludes this deployment.
    Limitations and uncertainty
    Faces seen are not established unique people. No causal baseline, demographic error analysis or September 4 alert adjudication is provided. HTML and PDF disagree on older entries; no totals or disputed figures are used. UK legal authority does not transfer to U.S. agencies.
    Government operator deployment monitoring, not causal evaluation Role takeaways & implementation
  4. Public-sector association guidanceEmergingNewly relevant · Mar 2026

    Judicial early-adopter interviews identify bounded uses and unresolved risks

    Judicial early adopters describe administrative and communication uses while retaining personal decision responsibility.

    TRI/NCSC AI Policy Consortium for Law & CourtsUnited States; state and federal judiciaryMarch 13, 2026

    Why it matters, evidence and limitations
    Why it matters
    New-to-archive U.S. court workflow context, not a newly issued September result.
    Evidence and measured results
    NCSC summarizes 13 one-hour interviews in 10 states during October–November 2025. Efficiency benefits are self-reported; concerns include hallucinations, privacy, deskilling and filing volume.
    Limitations and uncertainty
    Selected early adopters are not representative; no controlled outcome baseline. Only the substantive NCSC summary was accessible; linked full report returned a JavaScript shell. Do not infer interview detail or measured savings.
    Public-sector association substantive research summary Role takeaways & implementation

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.
Government evaluation
A public body’s measured evaluation or documented pilot.
Public-sector association guidance
Practitioner guidance or an association-supplied case; not independent outcome evidence.