From the Public Safety edition of September 20, 2026
CDT: Automated police-report drafting needs disclosure, audit trails, and use limits
Center for Democracy & Technology · Civil rights / technology policy (law enforcement) · United States
- Publisher
- Center for Democracy & Technology
- Original publication
- March 30, 2026
- Source retrieved
- 2026-09-21
What happened
CDT’s 30 Mar 2026 analysis of automated police-report drafting (BWC audio/video or field notes → generative narrative) argues reports shape charging, detention, pleas, and sentences, so AI drafts raise civil-rights stakes. It flags transcription errors (citing ~10% word-error rates even in ideal conditions), hallucination/speaker misattribution, automation bias/cognitive offloading marketed as a feature, and weak audit trails—noting some vendors advertise limited logs to “reduce FOIA exposure.” Recommendations: procurement transparency and independent accuracy verification; officer training/certification; restrict high-stakes use until empirical validation; mandatory disclosure plus retention of AI drafts/edit histories (citing CA SB 524 vs narrower UT SB 180); discovery rules covering AI drafts; independent bias audits. Newly relevant after 19 Sep Michigan Draft One / prosecutor-caution coverage: CDT supplies the civil-rights control checklist those pilots still need.
Why it matters
Why now: pairs with the stream’s 19 Sep Michigan Draft One and prosecutor-gate pattern—CDT was not previously archived and supplies procurement, discovery, and FOIA/accountability controls U.S. agencies should apply before expanding AI report writing.
Evidence and measured results
Policy analysis with cited examples (including bizarre Draft One frog hallucination anecdote) and statutory comparisons (CA SB 524; UT SB 180). Not a new randomized field trial; relies on cited technical literature for WER ~10% and known deployments (Axon, Truleo claims).
Limitations and uncertainty
March 2026 publication (older than last run) included for new relevance to 19 Sep Michigan wave; advocacy organization perspective; not a multi-site outcome evaluation.
Put this evidence to work
Lighthouse Advisory interpretation, grounded in this source. Enriched 2026-09-21; this does not change the original publication date. Each role follows the same checklist so takeaways compare across sources.
Sales
Role takeaway
- Customer problem
- Agencies want paperwork relief but face civil-rights, FOIA, and discovery blowback if AI drafts lack provenance.
- Stakeholders
- Police executive
- City/county attorney
- Prosecutor
- Civilian oversight
- Records/FOIA officer
- Discovery questions
- Does the vendor retain full edit history?
- Is AI use labeled for discovery?
- Are high-stakes offense types excluded?
- Who audits differential language impacts?
- Value hypothesis
- Bounded value is controlled draft-assist with reconstructible authorship—not FOIA exposure reduction.
- Potential engagement
- Procurement red-team and policy gap assessment against CDT checklist and state disclosure laws.
- Unsupported claims
- Guaranteed time savings
- That officer sign-off cures hallucinations
- Bias-free narrative generation
Pre-sales engineering
Role takeaway
- Fit
- BWC/field-note draft assist with mandatory provenance—not silent autofill of official narratives.
- Architecture and integration
- Source media ID → transcript → AI draft → tracked edits → labeled RMS narrative → retention store.
- Prerequisites
- CJIS/privacy review
- Edit-history capable product
- RMS AI flag
- Prosecutor input
- Constraints
- Vendors advertising FOIA-light logs
- Noisy BWC audio WER
- Mixed state statutes
- Security
- Contract bans on secondary use of audio; encryption; litigation holds on drafts/prompts.
- Proof of value
- Local WER benchmark
- Authorship reconstruction drill
- Prosecutor acceptance of discovery package
Delivery
Role takeaway
- Work
- Roll policy, training, labeling, retention job, and monthly supervisor accuracy sample.
- Dependencies
- Vendor audit features
- RMS configuration
- Training hours
- Union notice
- Ownership
- Patrol commander for field compliance; records manager for retention; legal for discovery.
- Skills and adoption
- Train officers to compare drafts to BWC before attesting personal knowledge.
- Governance checkpoints
- Pre-prod audit-log test
- 30-day sample review
- Expand/narrow offense types
- Acceptance criteria
- AI disclosure on filings
- Draft+edit history retained case-life
- FOIA-minimizing mode disabled
- Risks
- Automation bias
- Undisclosed AI text in charging packets
- Civil-rights disparate framing
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?
Require edit-history logging, AI-use labels in RMS, and source-to-draft linkage (BWC clip IDs); reject configurations that only store final sign-off.
Governance
Who approves, reviews and stays accountable for outcomes?
Treat FOIA-minimizing audit designs as a procurement red flag; align PD policy with prosecutor discovery expectations before rollout.
Security and privacy
What data, permissions and controls need testing?
BWC audio egress to vendors, prompt retention, and training-data reuse restrictions are central; limited logs impede both oversight and defense rights.
Accessibility and workforce
Who is affected, and what skills or accommodations follow?
Officers need recurring training to catch transcription/speaker errors; supervisors need rubrics that catch incompleteness, not just polish.
Procurement
What should contracts, pricing and exit terms secure?
Write RFPs requiring edit-history export, AI labeling, independent accuracy tests, and prohibition on marketing FOIA-avoidance as a feature.
Operating model
Which teams own the service once it runs?
Keep human attestation meaningful: policies should forbid filing AI text the officer has not compared to source evidence.
Publication history
- 2026-09-20Public Safety · Issue 144 resources
Stable resource ID: cdt-automated-police-report-drafting-20260330