New Zealand Police repeatedly broke their own rules while trialling AI transcription in criminal investigations — and a senior detective argued the force should keep using the tool anyway because restrictions “would likely hinder its effectiveness,” documents released under the Official Information Act have revealed. The OIA release was first detailed by IT Brief, with RNZ reporting the police responses in full.
The OIA files, reported by RNZ and published in full by IT Brief on Tuesday 20 October, show frontline investigators used the vetted transcription system Whisper AI outside its approved scope during a six-month test in 2024 and 2025, transcribing English audio when the rules limited it to translation. Officers also applied it to te reo Māori conversations, which breached the trial’s conditions, and turned to unsecured systems including Google Translate when the approved tool wasn’t available. Police leaders internally described some of the behaviour as “scary” and “risky workarounds.”
What the documents show
The trail of internal approvals is the alarming part. Police began using Whisper in 2024 “without having detailed auditing controls over it,” the reports show. By July 2025 — as the tool was being wired into the Cellebrite phone-harvesting platform investigators already use — the force’s security and privacy reference group wrote plainly:
“There are currently no SOPs [standard operating procedures], audit controls, or training materials governing its use, and its evidential limitations must be clearly communicated to staff.”
On public transparency, the same group reported: “No assessments or reviews are being considered for these new functionalities.”
Det Insp Stuart Mills of the National Criminal Investigations Group told RNZ there was “strong operational interest” in the technology. In a March 2025 report he argued officers should be allowed to keep using Whisper despite the repeated breaches, writing that “introducing additional controls to restrict the use of the service would likely hinder its effectiveness” and that automated transcription and translation tools were “essential” for investigations involving large volumes of foreign-language material — including the investigation into the killing of Shulai Wang, where fast translation of online communications sped up inquiries into what police describe as the Auckland cult killing.
Whisper was largely suspended in March 2025 because of the misuse. Police restarted it in October. In a statement on Monday, Mills said police are “continually assessing how they can use emerging technology to support investigations in a lawful and secure way consistent with their policies,” and that the force had “developed approved solutions” — without naming Whisper or any other tool. The force’s October 2025 Technology Capabilities List makes no mention of translation.
Barrister and privacy law expert Kathryn Dalziel told RNZ’s Checkpoint it was “absolutely unacceptable” that police took investigative data home and loaded it into open-source systems.
Why the safeguards lost to the workflow
The pattern in these documents is familiar from every other AI adoption story: the technology arrives through the side door, the safeguards are written after the fact, and the operational benefits are so immediate that the controls never catch up. A detective with a dozen hours of foreign-language audio has an obvious problem; Whisper solves it in minutes; the approval workflow takes longer than the investigation. The workaround becomes policy by practice long before anyone signs it off.
What makes this case sharper is the evidence chain. Translation errors in an investigation aren’t just inconvenient — they’re potentially miscarriage-of-justice material. An automated transcript that quietly mistranslates a key phrase, fed into a phone-harvesting platform without audit controls, gives defence lawyers and appeal courts exactly the kind of untraceable failure they look for. Police told RNZ they’re “assessing how to use emerging technology” after the report; the OIA documents show that assessment was supposed to have happened in 2024, before frontline staff needed it.
New Zealand’s Privacy Act requires that personal information be collected and used with appropriate security safeguards, and the Office of the Privacy Commissioner has issued specific guidance on using AI tools with personal information — guidance built precisely to stop the “paste it into a convenient tool first, ask permission later” sequence these documents describe. The OPC and the Privacy Commissioner could not be reached for comment on whether they have engaged with Police directly about the Whisper trial.
The NZ angle
This is a small, local story with national consequences. New Zealand police forces are not unusual — law enforcement agencies worldwide are wiring generative tools into investigations faster than their oversight bodies can write policy, and the US, UK and Australian equivalents have all had their own version of this scandal. But New Zealand has an unusually concentrated dependency: one national force, one privacy regulator, and a justice system small enough that a flawed evidentiary practice could propagate through dozens of prosecutions before anyone notices. The speed-versus-safeguards trade the detectives’ emails describe won’t be resolved by a single technology review — it will be resolved by whoever gets to define “essential” first.
As of publication, Police have not named which approved AI transcription tools have replaced Whisper, and the force says controls have tightened but the internal audits of the 2024–25 trial period remain internal. RNZ’s OIA documents are the only public record.
Further reading on singularity.kiwi: When AI Watches the Neighbourhood — the push to connect police, hospital and city data; China Regulated AI Agents With an Iron Fist — NZ Should Use Carrots Instead; AI Compliance New Zealand: Complete 2026 Guide for Businesses.
