Police restart approved AI after officers used unapproved models

According to Official Information Act documents released to RNZ, New Zealand Police trialled OpenAI's Whisper for translating and transcribing non-English audio for six months up to March 2025, then halted the project after officers used it to transcribe English audio despite that being outside the trial's restrictions. An internal document stated: 'Trial misuse demonstrated limitations of relying on behavioural controls alone.' Police restarted the trial in September 2025 with additional controls, citing 'strong operational demand' and concern that turning the tool off entirely 'increased risk of staff using unapproved alternatives.' The police high-tech crime group has now used the approved model, internally called TranScriptor, for roughly nine months.

The documents state Whisper is 45 percent inaccurate on Maori and Pacific-language audio and is barred from use on those languages. RNZ also cites a May 2024 Cornell University study finding Whisper hallucinates in about 1 percent of transcriptions, including fabricating racial commentary, violent rhetoric, and invented medical treatments. OpenAI told RNZ, 'Addressing hallucinations is an ongoing area of research,' and that 'speech recognition systems are not perfect and should be evaluated carefully for their intended use case,' adding that newer variants such as GPT-Realtime-Whisper reduce hallucinations and improve accuracy.

Police policy requires an audit of generative-AI use every six months, but police confirmed to RNZ that no formal audit of Whisper has been conducted since the original 2025 trial evaluation, even though system use is logged per-user for potential future audit. An internal police warning from March 2025 noted the stakes were high partly because 'there is no specific legal framework for the use of AI in New Zealand.' Abdur Razzaq of the Federation of Islamic Associations, whose community's audio was processed under the trial, told RNZ, 'We should have been consulted.' Output is not used directly as evidence; police policy requires disclosure to courts when a draft AI transcript informs a final human-reviewed version used evidentially.

The recurring lesson across public-sector AI deployments is that language and accuracy gaps concentrate on underrepresented languages and dialects, and that policy controls without independent, scheduled technical audits leave a real accountability gap, especially where outputs could influence investigations or evidence.

Whether police complete the overdue Whisper audit, whether the Ministry of Justice's parallel transcription-technology tender adopts different safeguards, and whether affected community groups are consulted on future generative-AI deployments in law enforcement.