Problem being that ML manifests the unconscious biases of its author -- in the form of its learning set being biased, in the form of not-spotting patterns that are obvious for human heuristics or in the form of training not capturing something the author sees as significant.

I recall a story recently about ML detecting risk for post-op inpatients. Human nurses know not to send people with Asthma home after an operation due to their high risk profile. Because the ML system was always ignoring Asthma patients as they were never queried, it assumed it was safe to send them home.

I would argue that that racially biased policing in ML is just an extension of racially biased data scientists in Ft Meade, or Silicon Valley or Wall St or where ever.

I mean... I'd rather we don't have that kind of policing at all. Obviously...

You now have both.