{"resourceId":"tasmania-police-draft-one-family-violence-evaluation-errors-complex-calls-20261022","versions":[{"version":"external-3c8292985e17b6db8a6721371bf16a3a61cfaaaa6a682eff19e1fc7ca269f719","resource":{"id":"tasmania-police-draft-one-family-violence-evaluation-errors-complex-calls-20261022","title":"Tasmania Police evaluation of Axon Draft One for family violence reports found writing time nearly halved, but the tool confused roles and occasionally missed threats in complex incidents","organization":"Tasmania Police (Australia); Axon (vendor)","sector":"Policing / AI report writing","geography":"Tasmania, Australia","publishedAt":"Fri 23 Oct 2026 5:39am AEDT (Oct. 22, 2026 U.S. Central), ABC News","publicationDate":"2026-10-22","eventDate":null,"sourceName":"ABC News (Australia)","sourceLabel":"AI tool from US manufacturer Axon drafts family violence reports for Tasmania Police","sourceUrl":"https://www.abc.net.au/news/2026-10-23/ai-assisted-police-report-writing-dv-domestic-violence/107297138","evidenceClass":"independent-reporting","outcomeClass":"mixed","topics":["knowledge-work","governance-procurement","accessibility-workforce"],"finding":"Right to Information documents show Tasmania Police evaluated Draft One in two phases in 2025: about 150 randomly selected historical family-violence body-camera recordings, then a several-month Launceston field test comparing human-written and AI-assisted reports. Reports were written and validated in 42 minutes, nearly twice as fast, which the evaluation projected as about 160 weeks of work time and $434,187 in wages a year. It also flagged some coercive behaviours officers missed. But Draft One occasionally confused names and roles and in a small share of cases missed a threat, assault or risk factor; performance dropped when partners, children or witnesses were present. One draft said a woman alleged prior threats when she had said there were none. The evaluation recommended a phased rollout once ethical-use policies exist; police say rollout remains under review pending funding, consultation and more trials.","sledRelevance":"One of the few government evaluations of AI report drafting with a field comparison and a named failure mode. U.S. agencies piloting Draft One can use its finding that accuracy falls in multi-party domestic incidents, which are common and legally sensitive calls.","evidence":"Opened the ABC article by Jano Gibson, which reports on the evaluation obtained under Right to Information and quotes it directly; the evaluation document itself was not opened.","architectureImplications":"Interpretation: drafts depend on body-camera audio quality; overlapping speakers caused at least one material error, so audio capture and speaker separation are part of the system boundary.","governanceImplications":"The evaluation itself made rollout conditional on ethical-use policies, mandatory human verification and supervisor oversight; a victim advocate proposed letting victim-survivors have a say on AI use.","securityPrivacyImplications":"Family-violence recordings are highly sensitive; the article does not describe data handling or retention of AI drafts.","caveats":"Secondary reporting of an evaluation not opened directly; sample sizes for phase two, error rates and methods for the time comparison are not given; the \"small percentage\" of missed risk factors is not quantified.","streamIds":["public-safety"],"roles":{"sales":"Sales framing: this is a rare government evaluation of AI report drafting that shows both sides. Writing time nearly halved, yet the tool confused roles and sometimes missed threats in complex family-violence calls. A credible conversation is about where drafting is safe to use and what review it needs, not blanket savings. A bounded engagement could design a call-type-specific pilot with error tracking. This source cannot support claims of guaranteed accuracy, realised wage savings or suitability for domestic-violence reports.","engineering":"Engineering framing: the failure mode is specific and testable. Errors rose with multiple speakers, and one came from overlapping audio. Pre-sales should check audio capture quality, speaker separation and whether the agency keeps the original AI draft for comparison. Proof of value should score drafts against footage for role attribution, omitted risk factors and invented statements, separately for single-party and multi-party incidents, because the evaluation found accuracy differed sharply between them.","delivery":"Delivery framing: a rollout would need policy before scale, as the evaluation itself recommended. Work includes defining which call types may use drafting, a supervisor review step, and a log of corrections. Dependencies are licence funding, extra program staff and stakeholder consultation, including victim advocates. The operational owner is the police service, with prosecutors consulted on court use. Acceptance should be measured error rates on sampled reports, not time saved alone."},"roleDetails":{"sales":{"customerProblem":"Officers spend heavy time on family-violence reports, but AI drafts can misattribute roles or omit risk factors in complex calls.","stakeholders":["Police executive and family-violence command","Frontline supervisors","Prosecutors","Victim-survivor advocates","Privacy and ethics leads"],"discoveryQuestions":["Which call types would be in scope?","Do you retain the original AI draft?","How are supervisors expected to verify drafts?","How will victims be informed of AI use?"],"valueHypothesis":"Drafting may cut writing time for straightforward single-party incidents while keeping complex calls fully human-reviewed or out of scope.","potentialEngagement":"Eight-to-twelve week call-type-specific drafting pilot with draft-versus-footage error scoring.","unsupportedClaims":["Guaranteed wage savings","Equal accuracy in multi-party incidents","That human review reliably catches every error","Court acceptability in other jurisdictions"]},"engineering":{"fit":"Agencies with body cameras evaluating AI report drafting for high-volume call types.","architectureIntegration":"Body-camera audio feeds a drafting service; drafts flow into the records system after officer edit and supervisor approval.","prerequisites":["Reliable body-camera audio","Records system integration","Draft retention policy","Defined in-scope call types"],"constraints":["Accuracy drops with multiple speakers","Australian court and RTI context","Evaluation document not public in full"],"security":"Restrict access to sensitive family-violence audio and drafts; log edits between AI draft and final report.","proofOfValue":["Proposed: score role attribution and omitted risk factors against footage","Proposed: compare single-party and multi-party incident accuracy separately"]},"delivery":{"work":"Define scope, review workflow and correction logging before any phased rollout.","dependencies":["Licence funding","Additional program staff","Ethical-use policy","Prosecutor and advocate consultation"],"ownership":"Police service operations command, with legal and family-violence leads","skillsAdoption":"Train officers and supervisors to verify drafts against footage, not just edit wording.","governanceCheckpoints":["Policy approval before rollout","Interim error review after first months","Re-test after any model or audio change"],"acceptanceCriteria":["Proposed: role or threat omission errors below an agreed threshold on sampled reports","Proposed: every AI-assisted report carries supervisor sign-off"],"risks":["Missed risk factors in domestic incidents","Automation bias in review","Court challenges to AI-assisted reports"]}},"retrievedAt":"2026-10-23T02:57:49Z","enrichedAt":"2026-10-23T03:02:58Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Time savings were the main projected benefit; the evaluation said additional staff would be needed to run the program.","procurementImplications":"Funding for Axon licences and program staff is unresolved; savings are projections from the trial, not realised budget reductions.","operatingModelImplications":"Human review and supervisor sign-off are assumed to catch errors; a digital rights advocate questioned whether officers can rely on AI-assisted reports in court.","sourceVerification":{"openedUrl":"https://www.abc.net.au/news/2026-10-23/ai-assisted-police-report-writing-dv-domestic-violence/107297138","referenceExcerpt":"The evaluation also found Draft One occasionally confused names and roles, and, in a small percentage of cases, missed a threat, assault or risk factor.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}