Policija je tiho ustanovila laboratorij za testiranje novih strojev.

New Zealand Police admitted on Monday that they repeatedly violated their own safety protocols while using generative artificial intelligence in criminal investigations during 2024 and 2025. Official documents released under the Official Information Act reveal that officers utilized unvetted systems and breached specific usage rules for approved software. Despite these failures, the force stated it continues to evaluate emerging technologies to support investigations in a lawful and secure manner. Detective Inspector Stuart Mills of the National Criminal Investigations Group noted that investigators frequently handle fast-moving cases involving large volumes of foreign-language material. He emphasized the need to balance rapid identification of critical information with privacy and security considerations. The disclosures highlight a period of rapid technological adoption marked by procedural gaps. Police began using the Whisper AI translation system in 2024 without detailed auditing controls. Although the tool was intended specifically for translation, frontline investigators used it to transcribe English audio, violating the terms of its six-month trial. Mills argued in a March 2025 report that restricting the service would hinder its effectiveness, describing automated transcription and translation tools as essential. Consequently, although Whisper was suspended in March due to misuse, police restarted the system in October. By July 2025, the force’s security and privacy reference group warned that standard operating procedures, audit controls, and training materials were absent. The group further stated that no assessments or reviews were being considered for new functionalities integrating Whisper into existing phone-harvesting tools like Cellebrite. Privacy experts have criticized the handling of sensitive data. Barrister Kathryn Dalziel described it as absolutely unacceptable that police took data home and loaded it into open-source systems. Senior police leaders themselves characterized some of these workarounds as scary and risky. In addition to Whisper, officers used unsecured platforms such as Google Translate. The documents indicate that the force developed approved solutions, though Mills did not specify which tools were included. The latest Technology Capabilities List from October 2025 makes no mention of translation capabilities. Parallel to these operational issues, police established an AI innovation lab around August 2024. The lab’s mandate was to test capabilities and oversee the implementation of AI across the organization. It began work by investigating other translation tools and met with Europol in September 2025 to examine international practices. Europol maintains a repository of forty AI tools for law enforcement. The lab has assisted in evaluating several new technologies, including a translation tool from Meta, the parent company of Facebook. An AI forensic data analyzer named BelkaGPT was also examined. Belkasoft, the US-based vendor, describes BelkaGPT as a groundbreaking offline AI assistant for digital forensics and cyber incident response. A police memo noted that international partners were consulted regarding the tool, and no concerns were raised. The lab also supported the implementation of an AI transcription application for video interviews conducted in police suites. Initially limited to English-language recordings, the transcripts require verification by a qualified person to ensure accuracy. A Public Service Commission report indicated that faster transcript availability would streamline operations and allow the public to be processed more promptly. The report also confirmed that the Serious Fraud Office has been utilizing Whisper AI to speed up the initial screening and triage of investigative material. The agency is developing additional internal tools, including predictive coding for large datasets and AI-enabled image categorization for forensic images. Plans for expanding AI integration continued through late 2025. By April 2026, police began considering the addition of Meta’s Seamless AI translation tool alongside Whisper. Internal correspondence suggested that a lighter risk assessment would suffice for Seamless compared to the eleven-page policy risk assessment performed for Whisper in 2024. A senior assurance advisor on information security advised the High Tech Crime Group in June that the risk profile would remain similar, recommending a one or two-page letter of assertion instead of a comprehensive review. This approach reflects a broader strategy outlined in internal memos, which aims to integrate AI-enabled forensic tools safely into existing workflows. The force cited the increasing volume of complex data from mobile devices, cloud platforms, and Internet of Things devices as a driver for early adoption. Police maintain that they adopt new technology based on operational need, risk, legal requirements, and policy considerations. They assert that robust testing and assurance processes are in place to maintain public trust and protect privacy. However, the released documents paint a picture of an organization struggling to align its rapid technical advancements with established regulatory frameworks. The tension between operational urgency and procedural compliance remains central to the ongoing debate over the role of AI in New Zealand law enforcement.