To each their own. Computer vision of this kind is not like cutting edge ai. This is pretty standard fare and has been for decades

Finger movements are generally on the order of 1-2Hz, a hand holding a large piece of wood will generally be much lower than that, which means that at the hundreds of millisecond level most of the movement can be predicted from momentum alone. Something which identifies and tracks hands in a view and fits a second order model to the movement can likely predict accurately enough at that timescale to make for a meaningful safety improvement (especially because if it's non-destructive you can tune it to err more on the side of caution than a destructive option)

Holding a piece of wood and sliding it along a table saw (to cut it) is the canonical method for losing a finger, and you could definitely pick that out with a relatively simple bit of computer vision.

two cameras