> A machine can be tested for bias and measures taken against such bias are certainly followed.

To what degree? Based on first-order data, some populations really do commit more crime. However, we know that background/poverty is the real, underlying cause.

And even if we fed enough data to an algorithm to avoid that bias, avoiding the bias isn't enough. What you really want to do is tackle the cause. If the AI is a black box, you'd never know.

The problem with AI is almost always a lack of transparency, especially in the specific case, and not the general/overall case. Most training doesn't focus on this. Besides, the AI would limited by how good the bias tests we can come up with are.

Edit: For some use-cases, this general case optimisation is a feature, e.g. token customer service for cheap companies. Also much easier to control people who aren't that different.