Not at all a similar question, the outcome of machine learning of driving a car can be judged on objective results (incidents, accidents, injuries, deaths, etc)

Sorry, can't agree. You can validate the performance of dogs in a test setup, just like you can validate machine learning in a sandbox environment.

You're right there but that's not why these are not similar questions.

A ML agent driving a car is about being able to do something out of "it's own agency"(We'll pass the free will observations for now, they're really irrelevant unless that's the point one makes between a ML agent and a dog, where the latter arguably has it compared to the former), the focus here being on "doing". A dog's sniff being considered as enough evidence to put someone in jail is not really the equivalent to "the dog has put X in jail".There's a case to resolve, and obviously the person/agent who has the evidence does not put the criminal in jail.Here the focus would be on the fact that the dog is not only observing something, but also the subjective interpretation of it's senses (assuming we would have such a thing as an "objective interpretation of one's senses at a given time", which is problematic imo)

If ML had well documented biases towards hitting people on purpose, then no.

That similar question and this entire thread reminds me of xkcd 2173. [0] We obviously already let neural networks detect drugs and drive cars, though the former gets to wear a nice fluffy fur coat.

[0] https://xkcd.com/2173/