From the Public Safety edition of September 27, 2026
Milwaukee PD considers generative AI report tools under January SOP amid Brady and hallucination warnings
Wisconsin Watch / Milwaukee Neighborhood News Service · Independent nonprofit journalism · Milwaukee, Wisconsin, United States
- Publisher
- Wisconsin Watch
- Original publication
- September 26, 2026
- Source retrieved
- 2026-09-28
What happened
On Sept. 26, 2026, Wisconsin Watch republished Milwaukee Neighborhood News Service reporting that the Milwaukee Police Department is not currently using generative AI but has a January 2026 generative-AI SOP and a Generative Artificial Intelligence Committee evaluating products such as Axon's Draft One, which drafts police reports from body-worn camera audio. An MPD spokesperson said onboarding remains in the consideration phase and the department is not anywhere close to implementing the technology. The SOP, opened as a linked PDF (issued/effective Jan. 23, 2026), requires treating AI output as draft, reviewing for accuracy, disclosing AI use, and states generative AI does not waive otherwise-applicable laws/policies. ACLU of Wisconsin's Amanda Merkwae warned of LLM errors, hallucinations, and training-data bias, and said conflicts between officer testimony and an AI-assisted report could raise credibility questions that put an officer on a prosecutor's Brady list. Milwaukee County Chief Deputy DA Jeffrey Altenburg said generative AI can help with human oversight; Deputy DA Matthew Torbenson stressed officers' on-scene sensory impressions (e.g., child-neglect home conditions) will not come from AI. Wisconsin State Public Defenders' Connor Romenesko warned that rushed end-of-shift failures to review AI drafts can still produce charging errors without malice.
Why it matters
U.S. municipal policing generative AI for report writing and criminal-case evidence quality—public-safety courts interface, not fire/EMS.
Evidence and measured results
Sept 26 nonprofit journalism with named MPD/DA/ACLU/public-defender quotes; corroborating Sept 21 NNS original and opened MPD SOP 682 PDF.
Limitations and uncertainty
No MPD measured pilot outcomes; Brady risk is advocate scenario; DA comments are not a formal ban like King County's (noted in separate trade survey inspection); Wisconsin Watch is a republication of Sept 21 NNS.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-09-28; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Agency is evaluating generative report tools under a new SOP while prosecutors and defenders already flag hallucination and Brady/credibility risks.
- Stakeholders
- Police chief / Generative AI Committee
- City attorney
- County district attorney
- Public defender leadership
- Civil-rights advocates
- Discovery questions
- Has the DA stated whether AI-assisted narratives will be accepted for charging?
- Will original AI first drafts and audit trails be retained for discovery?
- Which incident classes are in/out of any pilot?
- How will testimony conflicts with AI-assisted text be handled for Brady?
- Value hypothesis
- Bounded value is a prosecutor-defensible evaluation and disclosure design—not guaranteed report-time savings or crime outcomes.
- Potential engagement
- 4–6 week generative-report readiness: SOP-to-procedure map, disclosure/RMS fields, DA acceptance workshop, Brady tabletop, go/no-go brief.
- Unsupported claims
- That MPD has selected or deployed Draft One
- Guaranteed reduction in report-writing hours
- That human review eliminates hallucination risk
Pre-sales engineering
Role takeaway
- Fit
- BWC-equipped agencies considering LLM draft narratives under written generative-AI policy before prosecutor acceptance is settled.
- Architecture and integration
- BWC audio to vendor draft narrative to officer edit/attest to RMS disclosure plus first-draft archive to charging package export with AI flag.
- Prerequisites
- Operationalized SOP 682
- RMS fields for AI disclosure and draft retention
- CJIS-compliant vendor path
- Prompt/output audit logs
- Constraints
- DA expectation that sensory observations remain officer-authored
- Brady exposure on testimony/report conflicts
- Labor impact if review duties expand
- Security
- Treat prompts as sensitive (tactics/CI/PII); default deny cloud training on agency data; log every generation and edit.
- Proof of value
- Sample packages accepted by DA reviewers
- 100% AI disclosure on audited reports
- Documented mismatch rate vs BWC ground truth
Delivery
Role takeaway
- Work
- Stand up evaluation, map SOP to procedures, run dual-review QA, brief DA/PD, decide go/no-go before production.
- Dependencies
- Vendor contract/BWC stack
- RMS change window
- Training and bargaining if duties change
- Prosecutor written posture
- Ownership
- MPD tech/professional-standards lead with DA liaison co-owner for acceptance criteria.
- Skills and adoption
- Train officers that AI drafts are not sensory evidence; train supervisors to spot rubber-stamping.
- Governance checkpoints
- Committee evaluation memo
- DA acceptance letter or rejection
- Pilot metrics review before expansion
- Acceptance criteria
- Written prosecutor posture on AI narratives
- Disclosure plus first-draft retention live
- Published pilot error/mismatch metrics
- Risks
- Silent production use without DA buy-in
- Brady disputes from AI/testimony drift
- Over-trust of fluent but wrong narratives
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?
Any Draft One-class BWC-to-narrative pipeline needs immutable first-draft retention, disclosure flags in RMS, and human attestation workflows before charging packages leave the agency.
Governance
Who approves, reviews and stays accountable for outcomes?
A January SOP plus a standing AI committee is pre-procurement governance; prosecutors and defenders are already defining acceptance constraints (oversight, Brady risk) before go-live.
Security and privacy
What data, permissions and controls need testing?
SOP itself flags prompt leakage of tactics/CI/active investigations and biased/inaccurate outputs—controls must cover model inputs as well as narrative accuracy.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Officers need time and training to verify AI drafts; supervisors and prosecutors need capacity to detect rubber-stamping; defense needs discovery of AI provenance.
Procurement
What should contracts, pricing and exit terms secure?
Any Draft One-class purchase should contract for first-draft retention, audit exports, training-data restrictions, and prosecutor acceptance gates—not only list-price efficiency claims.
Operating model
Which teams own the service once it runs?
Keep generative report tools in evaluation until prosecutor/defender constraints are written into acceptance criteria; policy alone is not deployment readiness.
Publication history
- 2026-09-27Public Safety · Issue 214 resources
Stable resource ID: wisconsinwatch-mpd-generative-ai-20260926