In my experience with machine learning, the last few percent of weird cases are exponentially varied and change over time so are always new. You hit a performance wall at a certain level. Researchers have been driven mad by this wall. See for example https://www.wired.com/2008/01/ff-aimystery/

For cars, behaving dangerously just 0.1% of the time is way too much.

A good AI should have really good 'one shot' deduction capability not just work by induction using tons of data. I don't think the current crop of algorithms are there yet.