Danish police to transcribe interrogations with AI, building own language model

Danish police are developing AI systems to transcribe interrogations, and are building their own language model to do it. The article, published by the Danish tech outlet Ingeniøren on 12 October, says the National Police (Rigspolitiet) is running a number of AI projects that will play a central role in the justice system, but that this is not something the force talks loudly about from its Copenhagen headquarters.

The projects were found by digging into the National Police's 2025 annual report, which lists three of them. Only the first title, POLtekst, is visible in the available text, and the rest of the article is behind a paywall. The headline and standfirst say the police are building their own language model, and that an expert believes the National Police is wading into "shark-infested waters" by developing different AI systems meant to process sensitive personal data. The accompanying caption notes that National Police Chief Lasse Boje Nielsen already has plenty of data problems to deal with and may now get another.

Reader comments on the piece show some of the debate. One commenter asks why speech-to-text, which has existed for years and maps input to output fairly directly, is now being routed through a language model that may guess at what the "right" word is. Others discuss how interviewees should handle the result. Two suggest that the person questioned should read and sign the AI-written record, as with a police-written one, or be given the audio or video recording. Another points out that a person questioned has no duty to sign a record and can ask for changes if it is inaccurate. One commenter says local models exist that avoid reliance on big tech, and questions the purpose of the chat services.

The full details of the three projects and the expert's concerns are not available in the text provided.