From the Public Safety edition of September 19, 2026
Fair and Just Prosecution: AI-generated police reports pose accuracy, accountability, and Brady risks
Fair and Just Prosecution · Prosecution / criminal justice policy · United States (national prosecutor network; King County, WA case study)
- Publisher
- Fair and Just Prosecution
- Original publication
- June 2025
- Source retrieved
- 2026-09-20
What happened
FJP’s June 2025 Issues at a Glance brief argues AI-generated police reports present high-stakes risks: hallucination/inaccuracy from poor audio and overlapping speakers; cognitive offloading that dilutes officer ownership; bias amplification; privacy exposure via vendor processing of BWC audio; minimal proven time savings (citing Adams et al.); and legal exposure including Fourth Amendment fact errors, Brady demands for model details, and impeachment if officers certify flawed AI text. Case study: King County Prosecuting Attorney’s Office does not accept AI-assisted narratives and requires reports produced entirely by the authoring officer after seeing an AI report reference an officer not at the scene. Recommendations urge prosecutors to set non-acceptance or strict safeguard policies and demand independent validation before adoption. Why now: newly relevant to 2026-09 Michigan Draft One adoption and Wayne County prosecutor caution documented in this edition’s WXYZ source.
Why it matters
Independent prosecutor-side scrutiny framework that explains why DA acceptance gates—not vendor demos—should control AI report rollout in SLED agencies now piloting Draft One.
Evidence and measured results
Opened FJP PDF via WebFetch and curl/pdftotext; confirmed King County non-acceptance case study, risk categories, Adams time-savings citation, and recommendation that prosecutors decline or tightly govern AI reports.
Limitations and uncertainty
Policy brief (June 2025), not a new RCT; King County example is office policy anecdote; quantitative claims largely cite external studies; why-now depends on current Michigan adoption context rather than a 2025–26 update to the PDF itself.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-09-20; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Prosecutors need charging-ready police narratives while vendors push AI drafting that may introduce unreviewable errors and discovery burdens.
- Stakeholders
- District/State’s Attorney
- Police chief
- Defense bar
- Court clerks
- Vendor account team
- Discovery questions
- Does the DA accept AI-assisted narratives today?
- What Brady materials can the vendor produce?
- Have independent accuracy studies been run on local audio?
- How is AI assistance disclosed in discovery?
- Value hypothesis
- Bounded value is clearer acceptance rules and safer pilots—not proven efficiency or accuracy gains.
- Potential engagement
- Prosecutor–PD joint policy sprint defining non-acceptance or safeguard conditions before any expansion.
- Unsupported claims
- Vendor-claimed time savings as proven
- That officer review eliminates hallucination risk
- Universal fitness of Draft One for all case types
Pre-sales engineering
Role takeaway
- Fit
- Only as tightly controlled draft assist—or disabled—when prosecutor policy forbids AI-assisted narratives.
- Architecture and integration
- Optional draft service behind policy flag; force officer rewrite path; persist AI-use metadata for Brady; block silent AI finalize.
- Prerequisites
- DA written policy
- Vendor disclosure pack
- RMS AI-flag
- Audio quality thresholds
- Constraints
- Proprietary model opacity
- Brief is not a local empirical study
- Overlapping-speech/noise failure modes
- Security
- CJIS-aligned vendor controls; deletion SLAs; no secondary use of BWC audio for model training.
- Proof of value
- Independent accuracy evaluation
- Prosecutor challenge rate
- Impeachment incidents tied to AI text = 0
Delivery
Role takeaway
- Work
- Draft DA acceptance policy, train officers on ownership/attestation, and stand up audit sampling of AI-touched reports.
- Dependencies
- FJP-style policy models
- PD vendor contracts
- Discovery unit procedures
- Ownership
- Elected prosecutor owns acceptance standard; chief owns field compliance.
- Skills and adoption
- Reinforce that certifying a report means personal knowledge—not rubber-stamping AI prose.
- Governance checkpoints
- Policy issuance
- First-month audit
- Revisit after any AI-error incident
- Acceptance criteria
- Published DA rule on AI reports
- AI assistance detectable in case files
- Sampling finds no unedited Easter-egg/hallucinated content
- Risks
- Brady litigation over model secrecy
- Officer Brady-list exposure
- Erosion of public trust after AI errors
Implementation considerations
Lighthouse Advisory interpretation across the operating dimensions a public-sector buyer must settle before this evidence becomes a design. Each note answers the question under its heading for this specific source.
Architecture and integration
What must connect, and where does the AI sit in the workflow?
If used at all, isolate vendor LLM drafting from final RMS narrative authority; preserve prompts/versions for Brady; prefer prosecutor-reject flags in RMS.
Governance
Who approves, reviews and stays accountable for outcomes?
Prosecutors should publish acceptance/non-acceptance rules before PD purchase; require independent accuracy studies and disclosure protocols.
Security and privacy
What data, permissions and controls need testing?
BWC audio to vendors creates retention and misuse risks; contracts need CJIS, deletion, and no-train clauses.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Offices without training capacity for automation-bias resistance should not expand AI drafting; FJP stresses skill and accountability erosion risks.
Procurement
What should contracts, pricing and exit terms secure?
Require independent evaluations, Brady-producible documentation, and DA acceptance clauses before AI report tool purchase.
Operating model
Which teams own the service once it runs?
Make prosecutor acceptance the control point for AI police-report operating models.
Publication history
- 2026-09-19Public Safety · Issue 134 resources
Stable resource ID: fjp-ai-police-reports-risks-brief-202506