Public Safety & Emergency ServicesIssue 104 records
Public Safety
Four sources new to the archive: three September 2026 Council on Criminal Justice Assessing AI in Action case studies applying the User Decision Framework to police-report drafting (Careful Implementation; prompt retention as dealbreaker), state corrections classification (Title VI testing and contestability), and public-defender general-purpose AI (Standard Deploy only with confidentiality preconditions), plus August 2026 ACLU commentary on vendor incentives to shape police-report content. Findings emphasize governance gates over proven operational benefits. Two patterns address prompt/vendor opacity for report AI and Careful Implementation prerequisites for substantial-risk justice AI. No generalized U.S. savings or prison-outcome claims. Role guidance is interpretation.
- Evidence records
- 4
- Cross-source patterns
- 2
- Evidence classes
- 3 standards or public-body guidance1 independent reporting
- Outcomes
- 3 cautionary1 emerging
- Source freshness
- 3 undated1 recent
- Research completed
- 2026-09-17
Choose a role to see its takeaway beside every record in the ledger.
Patterns across the evidence
Synthesis · Lighthouse Advisory interpretation
Treat prompt opacity and vendor incentive conflicts as dealbreakers for police-report AI
The September 2026 CCJ police-report case study treats refusal to retain and produce system prompts as a Generally Avoid trigger, while the August 2026 ACLU commentary argues multi-product vendors may have incentives to shape what reports include. Together they support refusing or tightly conditioning report-drafting AI when prompts and drafts are not producible and conflict reviews are absent. Neither proves a named vendor currently distorts outputs, and neither measures net time savings.
Operating questionCan this agency produce prompts, initial drafts, and conflict disclosures for every AI-assisted police report before expanding use?
Supporting evidenceCCJ case study classifies AI police-report drafting as substantial risk requiring careful implementationAmerican Civil Liberties Union (Speech, Privacy, and Technology Project)
Substantial-risk justice AI needs Careful Implementation gates before scale
CCJ's September 2026 police-report and corrections-classification case studies both classify their focal uses as substantial risk plus substantial opportunity, requiring Careful Implementation—independent validation, enhanced rights protections, monitoring, and annual reassessment—rather than enthusiasm-led rollout. They do not prove operational benefits and remain hypothetical-agency guidance.
Operating questionWhich Careful Implementation requirements (validation, contestability, disparity monitoring, annual reassessment) are still open before this police or corrections AI use scales?
Supporting evidenceCCJ case study classifies AI police-report drafting as substantial risk requiring careful implementationCCJ corrections case study ties AI classification to Title VI testing, contestability and annual reassessment
Evidence ledger
Full record · every source keeps its link and limitations
CCJ case study classifies AI police-report drafting as substantial risk requiring careful implementation
September 2026 CCJ case study applies the March 2026 User Decision Framework to AI tools that draft incident narratives from body-worn camera audio or related inputs. For a hypothetical 200-officer mid-size department, it classifies the use as substantial risk and substantial opportunity, requiring Careful Implementation: independent recall before reviewing AI drafts, prompt/initial-draft retention for discovery beyond CA SB 524/Utah SB 180 minima, bias auditing, six-to-twelve-month pilots with pre-set go/no-go criteria, and annual reassessment. It cites independent RCT evidence contradicting vendor time-savings claims and treats vendor refusal to retain/produce prompts as a Generally Avoid trigger. It does not evaluate one named product in a live agency deployment.
Why it matters, evidence and limitations
- Why it matters
- Directly usable by U.S. police and prosecutors weighing Draft One-class tools: frames procurement, discovery, Sixth Amendment/Brady, automation-bias controls, and when to refuse vendors. Interpretation only; agencies must map to state disclosure statutes and local prosecutor acceptance.
- Evidence and measured results
- Opened original HTTPS HTML (WebFetch + curl). Case study explicitly labels Careful Implementation, documents hallucination examples (Heber City; King County), and cites Adams et al. 2024 RCT on time savings.
- Limitations and uncertainty
- Hypothetical agency framing drawing on practitioners/vendors/research; Microsoft among disclosed funders; not a controlled multi-site effectiveness trial; product performance varies by architecture.
CCJ corrections case study ties AI classification to Title VI testing, contestability and annual reassessment
September 2026 CCJ case study applies the User Decision Framework to AI-powered inmate classification supporting custody, housing, programming, and reclassification for a hypothetical 15k–25k person state DOC. It finds substantial risk and substantial opportunity, requiring Careful Implementation: Title VI-grounded disparate-impact testing as a contract term, case-specific decision rationales on contest, documented justification both when following and overriding scores, Community Advisory Committee with directly affected representation, independent validation, and annual reassessment. Prohibited-use screening passes only if disparate-impact testing and substantive contest can be negotiated; otherwise treat as prohibited until mitigated. Compares AI to existing arbitrary/under-explained baselines rather than ideal human judgment. Not a measured deployment evaluation.
Why it matters, evidence and limitations
- Why it matters
- Actionable for state DOC and local jail classification modernization: defines contract redlines (disaggregated accuracy, data portability, audit rights) and operational controls (override logging by demographic group). Interpretation; federal equal-protection intent doctrine alone is called inadequate—Title VI and state rules matter.
- Evidence and measured results
- Opened original HTTPS HTML. Screening table marks contest and disparate-impact as YES with mitigation required; risk section cites liberty impacts and documented disparities in existing tools (COMPAS/PATTERN references).
- Limitations and uncertainty
- Hypothetical DOC scenario; evidence of opportunity draws on structured classification literature generally, not a specific AI product RCT; proprietary tools vary widely in transparency.
CCJ defender case study allows general-purpose AI only when confidentiality preconditions enable Standard Deploy
September 2026 CCJ case study examines general-purpose AI for public defender research, drafting, discovery analysis, and case prep in a large urban office. Classification is low risk and substantial opportunity—triggering Standard Deploy—only if privacy/confidentiality preconditions are met: enterprise-tier or air-gapped tools, no-training-on-customer-data terms, MDM and network blocks against consumer AI for client work, and lawyer-in-the-loop verification. Without those preconditions, recommendation flips to Generally Avoid because consumer accounts create a privilege-loss pathway (citing Heppner). Notes ~300 judicial AI standing orders (as of March 2026) and large hallucination-case counts as operational context. Not an RCT of defender outcomes.
Why it matters, evidence and limitations
- Why it matters
- Useful for public defender offices, court IT, and county counsel setting acceptable-use policy: distinguishes efficiency opportunity from privilege risk. Cross-relevant to prosecutor offices with similar confidentiality duties. Interpretation; jurisdiction-specific disclosure matrices still required.
- Evidence and measured results
- Opened original HTTPS HTML. Executive summary and risk section state the binary privacy precondition; implementation details specify MDM, domain blocking, and Acceptable-Use Policy (Appendix J).
- Limitations and uncertainty
- Hypothetical office; opportunity claims reference secondary efficiency reports; criminal-defense-specific accuracy benchmarks remain limited; Heppner and standing-order counts are evolving.
ACLU argues police-report AI vendors may have incentives to shape what reports omit or emphasize
August 28, 2026 ACLU commentary (updated September 2, 2026) argues a structural conflict beyond prior ACLU opposition to LLM police-report drafting: vendors that sell both report-drafting AI and other policing products (ALPR, cameras, Tasers, real-time crime centers) may have incentives to influence whether and how those products appear in narratives. It recounts Vigilant/Motorola ALPR reporting controversies as analogy and highlights opacity of models, fine-tuning, and prompts—citing EFF analysis that Draft One is designed to defy transparency. Newly relevant alongside the September 2026 CCJ police-report case study because both center prompt/vendor opacity as an accountability failure mode. Advocacy analysis, not a controlled audit of a live Draft One deployment.
Why it matters, evidence and limitations
- Why it matters
- Gives procurement and oversight stakeholders an independent scrutiny lens: evaluate vendor product portfolios and prompt opacity as conflict risks when buying report AI. Complements CCJ's Generally Avoid trigger on prompt non-production. Interpretation; does not prove a specific vendor currently distorts outputs.
- Evidence and measured results
- Opened original HTTPS HTML; article:published_time 2026-08-28; modified 2026-09-02. Core thesis stated in opening paragraphs with historical ALPR analogy and EFF citation.
- Limitations and uncertainty
- Advocacy commentary; ALPR analogy is historical and not direct proof of current report-AI distortion; Axon-specific design claims rely partly on secondary EFF reporting.
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.
- Standards or public-body guidance
- Normative or advisory guidance from a standards body or public institution.
- Independent reporting
- Independent reporting with attributable sources but without a formal evaluation design.