Fascinating---most of the examples they showed were immediately legible to me, and not at all examples of "bad handwriting".
There's also plenty of edge cases: letters get wet, they smudge, get torn or folded, and so forth. A postal system that processes 420 million letters a day[1] with a 99% true positive rate on OCR still needs to divert 4.2 million letters for manual review.
This is a common pattern, and why I think AI risk is overstated…unless we suddenly decide to start wholesale writing off the consequences of tail risk.
In other words, AI doesn't have to do anything perfectly to replace a large percentage of the workforce.