Police officers and instructors in an Umeå trial give a HoloLens decision aid high marks

Thirty-nine is a small denominator, and it pools serving officers with instructors ^[3]. All of them ran one scenario, a high-risk vehicle stop staged as a tactical exercise, with the headset as their support tool ^[3].

The measured outcome was attitude. Participants reported a very positive view of the system's usefulness, ease of use and trustworthiness ^[6]. Those are judgements about a tool. Whether officers wearing it made better calls at the car, or kept the stop safer, is a different question, and the phys.org write-up does not describe a group that ran the same stop without a headset for comparison. It reports that the results show "significant potential for integrating AI support into police interventions" ^[13].

The order of events complicates even the attitude data. Before the exercise, participants attended two workshops, one on the principles of trustworthy AI and one giving hands-on time with mixed reality and the HoloLens ^[5]. That is sensible preparation for an unfamiliar device. It also means the people scoring the system's trustworthiness had just been taught what trustworthiness should look like. A high score from that group fits a good tool, and it fits an interested audience fresh from training just as well. "The results were better than I expected. The police officers were very interested in the technology and saw many possibilities," said Anders Schogster, a police officer and teacher at Umeå's Unit of Police Work ^[7].

The firmer finding is the risk analysis, because it rests on an assessment of the system's functions and does not depend on how anyone felt during one exercise. The researchers placed those functions on three risk levels, from low to high ^[8]. Ethical risks around privacy, transparency and algorithmic bias grew as the system became more automated ^[8]. Data storage is also on their list of risks ^[11]. "It's important that this does not become fully autonomous. It should support police officers in their decision-making, but the decisions must ultimately be made by humans," said Jonas Hansson, a researcher at the same unit ^[9].

I think the identification feature is where that rule will be tested first. The headset can be used to flag potential risks in an officer's field of view ^[2]. Deciding that something at a stop is a threat is exactly the decision Hansson wants left to a person ^[9].