{"resourceId":"wxyz-michigan-draft-one-prosecutor-caution-20260916","versions":[{"version":"external-8a13217d643ab286ff7c9f005bf8721fb3f56e5ae44ee9251c7c5b63fd33aa71","resource":{"id":"wxyz-michigan-draft-one-prosecutor-caution-20260916","title":"WXYZ: Michigan Draft One pilots show mixed results as Wayne County prosecutor urges caution","organization":"WXYZ 7 News Detroit","sector":"Law enforcement / prosecution","geography":"United States (Michigan: Warren, Dearborn Heights, Kent, Wayne, Macomb, Oakland counties)","publishedAt":"September 16, 2026 (updated September 17, 2026)","publicationDate":"2026-09-16","eventDate":null,"sourceName":"WXYZ","sourceLabel":"Local investigative television report","sourceUrl":"https://www.wxyz.com/news/local-news/investigations/michigan-officers-turn-to-ai-for-police-reports-wayne-co-prosecutor-warns-be-cautious","evidenceClass":"independent-reporting","outcomeClass":"mixed","topics":["knowledge-work","governance-procurement","accessibility-workforce","operating-model"],"finding":"WXYZ reports Michigan agencies adopting Axon Draft One (body-cam audio → AI narrative). Warren is ~3 weeks into a 60-day purchase pilot with 'mixed' early results: skilled writers see little benefit; weaker writers report large gains. Warren requires officers to write ≥10% of each narrative, remove deliberate 'Easter egg' errors, and attest accuracy. Dearborn Heights allows optional use; Kent County limits to non-violent offenses and says prosecutors welcome thoroughness. Wayne County Prosecutor Kym Worthy sent chiefs a letter stating AI must never be used without thorough authenticity checks and that AI-generated police reports are 'fraught with problems' that can jeopardize prosecutions; she is drafting a countywide policy. Piece cites USC research (Ian Adams) finding no time savings and lower accuracy ratings for AI-assisted reports, and notes King County (WA) non-acceptance and Connecticut pause.","sledRelevance":"Current U.S. police-report AI implementation signal with prosecutor-side acceptance risk—core public-safety operating-model issue for chiefs and DAs after prior archive coverage of academic Draft One trials.","evidence":"Opened WXYZ HTTPS article via WebFetch and curl; confirmed Worthy letter language, Warren 10% rule and Easter eggs, mixed pilot quote, Kent non-violent limit, and USC study citations. Publication meta: 2026-09-16, modified 2026-09-17.","architectureImplications":"Keep Draft One as draft-assist with forced edit/attestation gates; flag AI-assisted reports in RMS for prosecutor disclosure; avoid mandatory use on violent/complex cases until accuracy audits exist.","governanceImplications":"Coordinate PD policy with prosecutor acceptance rules before purchase; document Easter-egg tests, supervisor review, and AI labeling.","securityPrivacyImplications":"Body-cam audio sent to vendor LLM pipelines needs CJIS/privacy review, retention limits, and discovery planning for prompts/models.","caveats":"Early Warren pilot (~3 weeks), not an RCT; time-savings/accuracy claims rely on cited USC studies rather than Michigan measured outcomes in-article; Axon did not comment.","streamIds":["public-safety"],"roles":{"sales":"Sales framing: Michigan’s Draft One wave is a mixed implementation story—Warren’s mid-pilot results diverge by writer skill, while Wayne County’s prosecutor warns AI-authored reports can jeopardize cases and is writing countywide policy. Sell governance and prosecutor-alignment work, not guaranteed time savings. Stakeholders: chiefs, sheriffs, county prosecutors, unions, and IT. Bounded engagement: pilot design with acceptance criteria shared with the DA before purchase. Unsupported here: proven Michigan time savings or accuracy gains.","engineering":"Engineering framing: Integrate Draft One as optional narrative assist with AI-use flags, minimum officer-authored share, Easter-egg or adversarial error tests, and attestation before RMS finalize. Prerequisites: CJIS/vendor review, prosecutor policy coordination, and supervisor QA sampling. Constraints: Warren allows any crime type while Kent limits to non-violent—policy must be explicit. Security: control body-cam audio egress and discovery of system prompts. Proof of value: edit rates, prosecutor rejection rates, and independent accuracy audits—not vendor speed demos.","delivery":"Delivery framing: Deliver a time-boxed pilot with prosecutor liaison, officer training on verification, and go/no-go metrics. Dependencies: Axon licensing, BWC audio quality, and DA policy. Ownership: patrol command for use; prosecutor for charging acceptance. Skills: automation-bias resistance and narrative ownership. Acceptance: AI-assisted reports labeled; ≥agreed officer edit share; no unreviewed AI text filed. Risks: credibility hits if Easter eggs or hallucinations reach charging packets."},"roleDetails":{"sales":{"customerProblem":"Patrol leadership wants paperwork relief while prosecutors fear AI-generated narratives will jeopardize cases and officer credibility.","stakeholders":["Police chief / sheriff","County prosecutor","Patrol supervisors","Police union","RMS/IT"],"discoveryQuestions":["Will the prosecutor accept AI-assisted narratives?","What minimum officer-authored share and attestation are required?","How are AI-assisted reports labeled in discovery?","What independent accuracy metrics gate purchase?"],"valueHypothesis":"Bounded value is controlled draft-assist under prosecutor-aligned rules—not proven shift-wide time savings.","potentialEngagement":"60-day joint PD–prosecutor pilot charter with acceptance criteria and go/no-go before enterprise purchase.","unsupportedClaims":["Michigan-measured time savings","Accuracy superiority of AI drafts","That Easter eggs alone prevent courtroom risk"]},"engineering":{"fit":"Optional BWC-to-narrative draft assist inside report workflow—not autonomous report filing.","architectureIntegration":"BWC audio → vendor draft → officer edit UI with AI flag → supervisor sample → RMS; store AI-assisted indicator for discovery.","prerequisites":["CJIS/privacy review","Prosecutor written position","Attestation + minimum edit controls","QA sampling plan"],"constraints":["Early mixed pilot evidence","Crime-type policy variance across agencies","Vendor non-response in reporting"],"security":"Limit audio/PII egress; contractual bans on training reuse; preserve prompts/versions for litigation holds.","proofOfValue":["Prosecutor rejection/challenge rate","Officer edit-distance metrics","Blind accuracy review vs officer-only baselines"]},"delivery":{"work":"Stand up pilot cohort, training on Easter-egg/error hunting, supervisor QA, and weekly prosecutor sync.","dependencies":["Axon Draft One license","BWC platform","RMS AI-flag field","DA policy draft"],"ownership":"Patrol commander owns field use; prosecutor owns charging acceptance standards.","skillsAdoption":"Train officers that attestation means personal knowledge—AI text is not a substitute for recollection.","governanceCheckpoints":["Pre-pilot DA letter","30-day accuracy sample","Purchase go/no-go"],"acceptanceCriteria":["AI assistance labeled on every filing","Documented officer edits meet policy minimum","Zero known Easter eggs in filed reports"],"risks":["Automation bias on weaker writers","Brady/impeachment exposure from AI errors","Skill atrophy in report writing"]}},"retrievedAt":"2026-09-20T03:10:34Z","enrichedAt":"2026-09-20T03:12:30Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Mixed skill effects imply training and supervision burden; weaker writers may over-rely on AI while skilled writers gain little.","procurementImplications":"Do not sole-source on vendor time-savings claims; require prosecutor acceptance language, AI labeling, and exit rights if accuracy gates fail.","operatingModelImplications":"Treat AI report drafting as a prosecutor-facing process change, not only a patrol efficiency tool.","sourceVerification":{"openedUrl":"https://www.wxyz.com/news/local-news/investigations/michigan-officers-turn-to-ai-for-police-reports-wayne-co-prosecutor-warns-be-cautious","referenceExcerpt":"under no circumstances should AI be used without thoroughly checking and double-checking its authenticity","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}