Police use two AI translation-transcription tools to reduce errors

Police say they are using two AI-powered translation tools because this cuts down the chance of errors.

RNZ revealed this week police set up an AI lab a year ago at the same time they were grappling with officers breaching AI safeguards and exposing sensitive information.

Police said the lab set up last October ran tests of generative AI in a secure isolated sandbox that did not involve live cases or public-facing work.

They have this year added a translator from the US company Meta, called Seamless, alongside another called Whisper AI used since 2024.

Having two allowed investigators to use the model that was most accurate for a particular language and reduced errors, they said.

The 'TransScriptor' platform the tools run on had an audit function to track who used it for what and no data was transmitted outside of the organisation, police added.

Internal reports and emails have revealed a struggle between investigators and technology overseers over deploying Whisper over the last two years. They agreed it was a big help to assess masses of digitally gathered evidence for later human checking, but debated the controls over it.

RNZ repeatedly asked police for an interview. They have instead issued statements on Monday - which did not mention AI - and Thursday afternoon.

Police have previously said "policing with consent" was vital to their adopting crime-fighting technologies.

Police in Thursday's statement said the AI Lab was set up to develop and test proof of concept AI solutions for police use cases, and explore the potential of Large Language Models (LLMs) and Generative AI "that could potentially enhance policing outcomes and operational efficiency".

It ran testing only in a secure, isolated, sandbox and only historic police data was utilised.

"It was secured to handle up to restricted information and no public notification was required."

They listed various governance documents that applied to AI.

In addition, Detective Inspector Stu Mills, acting director of the National Criminal Investigation Group said in another statement Meta Seamless was added around May this year to transcription and translation services.

Users securely submitted audio files through a web-based interface to the TransScriptor platform that used Seamless and Whisper AI, and then returned the done job to them.

"All processing occurs internally and no data is transmitted outside of the organisation."

TransScriptor included an audit function that provided visibility of system use, including who submitted requests, the length and language of recordings and the number of jobs processed, Mill said.

Providing users with a choice of translation models reduced reliance on a single AI engine and allowed investigators to select the model that produced the most accurate results for a particular language, dialect, recording quality or operational use case, he said.

"This helps mitigate the risk of translation or transcription errors that may arise when a single model performs poorly on specific languages or audio conditions."

He said the tools would be added to the police Technology Capabilities List when it was updated this month.

As for a forensic data AI analyser police looked into last year, called Belksoft X, including BelkaGPT, while it offered enhanced search, analysis, triage and data exploration feature, approval to trial the software had not been granted.