How is anything you just described different, in any way, from the existing situation in risk assessment and fairness in insurance? IE, everything we say about ML, the insurance business has already been doing for a century using statistics. Their models exist for two things: to increase the likelihood of profit (due to properly determining rates based on risk) and staying compliant (if there are laws that ensure some level of fairness).

Frankly I don't see any real difference from long-established principles in the insurance space.