Study: AI fails to reproduce human judgements about rule violations

Failing to reproduce might be a good thing. For example, categorizing dogs as aggressive on its own vs for use in a policy decision introduces all manner of human biases- a dog owner who is thinking about whether that policy would apply to themself would not be impartial. Human idiosyncrasies and biases lead to inconsistent judgements, and one of the things we'd like to have in judgements is consistency, at least. The harshness of an appropriate response to a judgement is, for the moment, still a human input to the system, as humans decide the policies.