{"resourceId":"nz-police-ai-innovation-lab-seamless-light-risk-assessment-belkagpt-20261021","versions":[{"version":"external-9f8d461cee7c14f1278f3c53a7ddc3e6c823f7177fbf86f41effb5e6a821775c","resource":{"id":"nz-police-ai-innovation-lab-seamless-light-risk-assessment-belkagpt-20261021","title":"NZ Police ran an AI innovation lab for over a year, testing Meta Seamless translation and BelkaGPT forensics, and planned only a \"light\" risk assessment for Seamless after an 11-page review of Whisper","organization":"New Zealand Police","sector":"Policing / investigations and digital forensics","geography":"New Zealand","publishedAt":"October 22, 2026, 6:23am NZDT (RNZ; October 21 in U.S. time)","publicationDate":"2026-10-21","eventDate":null,"sourceName":"RNZ","sourceLabel":"Police quietly set up AI lab to test new machines","sourceUrl":"https://www.rnz.co.nz/news/science-and-technology/2069792/police-quietly-set-up-ai-lab-to-test-new-machines","evidenceClass":"independent-reporting","outcomeClass":"cautionary","topics":["governance-procurement","data-security","knowledge-work"],"finding":"Official Information Act documents show NZ Police set up an AI innovation lab around August 2025 to test capabilities and track AI implementation. It met Europol, helped with Meta’s Seamless translator, the BelkaGPT offline forensic assistant and an AI transcription app for interview recordings verified by a qualified person. Police had done an 11-page risk assessment for Whisper in 2024; for Seamless, a security adviser said a one-to-two page \"letter of assertion\" should suffice. Whisper remains banned for Māori and Pacific languages. Whisper and translation tools are absent from the police public technology list.","sledRelevance":"This follows the archived report of officers breaking AI usage rules and adds the governance response: a lab, but lighter assurance for a replacement model and an undisclosed tool list. It is a concrete case of model-swap risk review being shortened, relevant to any agency approving one AI tool and later substituting another.","evidence":"Opened the RNZ article dated 22 October 2026 NZT and read the OIA findings on the lab, Seamless, BelkaGPT, the transcription app, the 2024 Whisper risk assessment and the language restrictions.","architectureImplications":"Tools range from offline forensic assistants (BelkaGPT) to translation and transcription models; swapping models within the same workflow was treated as low-change by police assurance staff.","governanceImplications":"The 2024 Whisper assessment said proportionality should be reconsidered when access widened; RNZ reports it is not known whether that occurred. Seamless was to receive a short letter of assertion.","securityPrivacyImplications":"Prior breaches involved evidence uploaded to unapproved services; the transcription tool is not on the public list.","caveats":"Based on OIA documents as reported; police declined interviews; whether the BelkaGPT trial occurred is unconfirmed.","streamIds":["public-safety"],"roles":{"sales":"Sales framing: NZ Police show that approving one AI model does not settle the risk of the next one. Substituting a translation model drew a much shorter review than the original. A bounded engagement is a change-control review for AI tools already approved, covering model swaps and expanded access. This source cannot support claims that Seamless or BelkaGPT are unsafe, or that a light assessment led to harm.","engineering":"Engineering framing: treat a model swap as a new deployment where accuracy, language coverage and data flows can change. Pre-sales should ask how an agency tracks model versions, whether replacement models are tested on the same languages and data, and whether offline tools stay offline. Proof should re-run the original acceptance tests on any replacement model before it reaches investigators.","delivery":"Delivery framing: the work is a change-control process for AI tools: criteria for when a change needs full reassessment, re-testing on representative data and updates to public disclosures. Dependencies are an AI inventory and assurance staff capacity. The owner is the technology assurance function. Acceptance should include documented reassessment triggers and test results for each model change."},"roleDetails":{"sales":{"customerProblem":"Agencies approve AI tools once, then swap models or widen access with lighter review.","stakeholders":["Technology assurance teams","Investigations and forensics leads","Privacy officers","Community liaison for affected language groups"],"discoveryQuestions":["What triggers reassessment of an approved AI tool?","How are model replacements tested?","Which tools are omitted from public lists and why?","Which languages are excluded?"],"valueHypothesis":"Clear change-control rules can prevent model swaps from bypassing the scrutiny the original tool received.","potentialEngagement":"Change-control review of approved AI tools with reassessment criteria.","unsupportedClaims":["That Seamless is less accurate than Whisper","That the lab improved outcomes","That harm resulted"]},"engineering":{"fit":"Agencies using translation, transcription or forensic AI tools in investigations.","architectureIntegration":"Version tracking for models inside approved workflows, with re-test gates.","prerequisites":["Model inventory","Original acceptance tests","Representative multilingual test data"],"constraints":["Low-resource language accuracy","Vendor model updates outside agency control","New Zealand legal context"],"security":"Verify offline tools stay offline and evidence never reaches unapproved services.","proofOfValue":["Proposed: re-run acceptance tests on replacement model","Proposed: confirm language coverage before release"]},"delivery":{"work":"Write reassessment triggers, re-test model changes and update disclosure lists.","dependencies":["AI inventory","Assurance capacity","Vendor change notices"],"ownership":"Technology assurance function","skillsAdoption":"Brief investigators on why model changes require re-testing.","governanceCheckpoints":["Reassessment decision recorded for each change","Public list reviewed after changes"],"acceptanceCriteria":["Proposed: every model change has a recorded reassessment decision","Proposed: test results on file before rollout"],"risks":["Investigator pressure to skip controls","Undisclosed tools","Language bias"]}},"retrievedAt":"2026-10-22T02:51:53Z","enrichedAt":"2026-10-22T02:56:39Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Language restrictions remain for te reo Māori and Pacific languages; translation tools affect minority communities disproportionately per the risk assessment.","procurementImplications":"BelkaGPT vendor vetting relied on consulting international partners, with no concerns raised.","operatingModelImplications":"A central lab now evaluates AI tools, while investigators pressed against Technology Assurance controls.","sourceVerification":{"openedUrl":"https://www.rnz.co.nz/news/science-and-technology/2069792/police-quietly-set-up-ai-lab-to-test-new-machines","referenceExcerpt":"NZ Police have been running a lab to test artificial intelligence (AI) applications for more than a year.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}