- On October 19, 2026, RNZ published documents showing that New Zealand Police repeatedly breached AI usage rules when transcribing and translating material, creating risks of confidential investigative information being exposed.
- During a 2024 murder investigation in Auckland, police used Whisper AI to process 250 hours of Chinese-language recordings from more than 2,500 files. The system completed the task in 10 hours, compared with 750 hours and an estimated cost of US$50,000 for manual processing.
- AI helped identify important conversations, but police still required qualified interpreters to verify all evidence.
- Following a six-month trial, police identified multiple cases of Whisper being used contrary to policy, including for English transcription and the preparation of documents for legal proceedings.
- In March 2025, police stopped deploying Whisper widely, allowing exceptions only through a multi-level approval process.
- Under pressure to process evidence, investigators turned to personal devices and unsecured platforms. In one case, 350,000 lines of Chinese-language data were uploaded to Google Translate to help classify information.
- Police leaders warned that the lack of approved internal AI tools would continue to push staff toward unsafe alternatives.
- By October 2025, Whisper had been integrated into Cellebrite for investigative use. However, the documents did not clarify the auditing and monitoring arrangements for its use over the following 12 months.
📌 The incident highlights the tension between security requirements and pressure to improve investigative efficiency. Whisper reduced the processing time for 250 hours of recordings from approximately 750 hours to just 10 hours. However, the lack of an official deployment framework drove investigators toward insecure platforms, including uploading 350,000 lines of data to Google Translate. The documents underscore the need to develop secure internal AI capabilities supported by clear approval, verification, and monitoring procedures.
