Regarding your first two points, a proper evaluation would involve some estimate the relative utilities of those two outcomes, and an estimate of their frequenc

Regarding your first two points, a proper evaluation would involve some estimate the relative utilities of those two outcomes, and an estimate of their frequencies. Taking a cue from signal detection theory, one would want to design the algorithm such that the true positive rate maximizes expected utility, and takes into account how the probabilities shift in a situation. A tough task, on all accounts; however, the situation as it stands now is no different, except that, perhaps, we have less control over how we adjust the decision threshold (one thing that is apparent here is that there is a potential conflict between the individual utility of the police officer and the sort of social welfare function that the role of the police is supposed to fulfill). With respect to the diffusion of responsibility effect, I think you are probably right that there are lots of possible unintended consequences. In as much as an officer decision to pull the trigger explicitly or implicitly takes in to account the reduced feelings of guilt and increased feelings of peer support in case the weapon is in fact fired, and if it takes into account the smaller likelihood that pulling a trigger will actually discharge the weapon, it may well increase the frequency which the trigger is pulled. More darkly, one could imagine a squad of officers making it into a distributed game of Russian Roulette. On the other hand, I think this highlights the limits of our imagination. For example, one could imagine the following alternative. Each officer has a body cam, and a pair at base, who is actively monitoring the camera. A decision to fire the weapon must involve both the person on the ground, and the buddy. Or one could have a supervisor of some sort, who can override individual decisions, and so forth. I don't have the answer, but I think it warrants thinking about.