By Hamza Faisal
New Zealand police investigators repeatedly misused artificial intelligence translation tools in "seriously risky" ways, including uploading 350,000 lines of intercepted Chinese audio to Google Translate — a platform police itself considers unsecured for evidence — newly released official documents show.
The revelations come from hundreds of pages of police emails and internal reports obtained by RNZ under New Zealand's Official Information Act. They paint a picture of an overstretched force turning to AI shortcuts under crushing investigative workloads, and struggling to keep those shortcuts inside the rules.
The trouble began after an early success. In Operation Parade — the 2024 Auckland investigation into the killing of Shulai Wang, whose body was found in a bag in Gulf Harbour in July 2024 — detectives used OpenAI's Whisper system to work through 250 hours of intercepted Mandarin-language calls spread across more than 2,500 files. Whisper did in 10 hours what police estimated would have taken human interpreters 750 hours and cost $50,000, flagging the conversations worth prioritising. All evidence was later validated by qualified human translators.
Buoyed by the result, automated translation spread to other teams — and so did the rule-breaking.
Guards broken, rules ignored
Police set up a formal six-month test of Whisper with clear controls: no transcribing English audio, no use on Māori or Pacific languages. An evaluation found officers had broken the no-English rule many times. In one case, Waikato investigators went further, using AI to generate a document intended for disclosure in legal proceedings — a practice banned under every version of the rules.
The force's own High Tech Crime Group warned early on that "the ongoing compliance is beyond current resourcing" — in other words, nobody had the staff to police the police's AI use. Investigators themselves described the approval process as exhausting. "Man what a process trying to get this approved!" one Western Bay of Plenty detective wrote in July 2025, while a technology manager protested the "constant change" in the rules, adding: "We are exhausted."
Some officers resorted to taking evidence home to process on personal devices.
The Google Translate workaround
After Whisper was dropped in March 2025 — kept only for rare cases needing triple sign-off — the workarounds got riskier. In one 2025 financial investigation, officers scraped phones for WeChat threads using the Cellebrite phone-forensics tool, then fed 350,000 lines of Chinese audio into Google Translate, arguing they needed it to grade the material in time.
"Without being [able] to grade the data in a timely manner this investigation will have the potential of running for years damaging the police's reputation (trust and confidence, etc)," the officers emailed.
A police report acknowledged the uploads risked "exposing confidential information."
"Unfortunate but not unexpected"
Detective Inspector Stu Mills, the force's Intercept and Technology Operations manager, described the breaches as "unfortunate but not unexpected" under the pressure investigators faced. He warned that without a permanent, approved AI capability, staff would keep drifting toward unsecured services.
That prediction shaped the response: Whisper was re-introduced about a year ago and, by October 2025, was being folded into the Cellebrite workflows investigators already used. What remains murky is the oversight. The newly released papers shed little light on Whisper's use over the past 12 months — and it is unclear how, or even whether, its use is audited or the rules enforced.
By Hamza Faisal


