RNZ reports that police use of AI has restarted after officers used unapproved models. The same report says one of the systems was 45 percent inaccurate on Maori and Pacific languages, but was still used by officers. That detail should stop the conversation from becoming a generic debate about innovation. This is about power, language and trust.
Internal approval is not enough
Police work already carries coercive authority. When AI enters that environment, even a “low-risk” tool can shape decisions, priorities and confidence. A transcript tool that mishandles Maori or Pacific languages is not just technically flawed. It can distort evidence, misunderstand people, and deepen the suspicion that new technology works worst for the communities most likely to be scrutinised.
Restarting AI use may be defensible if the controls are strong. But the threshold cannot be “we fixed the internal process”. Public institutions need public legitimacy. People should know what types of AI are being used, what they are not allowed to do, how accuracy is tested, how errors are corrected, and who is accountable when a tool causes harm.
The standard should be simple
- No operational use without a documented purpose and human accountability.
- No language or transcription tool without independent testing across the communities it affects.
- No quiet expansion from administrative support into investigative or enforcement decision-making.
- No secrecy that prevents people from challenging AI-influenced errors.
New Zealand does not need to reject AI in policing outright. Some tools may save time, improve search, or reduce paperwork. But saving time is not the highest value in justice. Accuracy, fairness and contestability matter more. If an AI system helps police move faster in the wrong direction, that is not efficiency. It is risk at scale.
The better path is slower and cleaner: publish clear categories of use, commission independent audits, involve Maori and Pacific language expertise, and create a public reporting cycle. Police AI should not return as a quiet administrative restart. It should return only when the public can see the rules.
