If your measurement system is not in statistical control, then the first thing to do is to get it into statistical control, so that the measurements that it takes are suitable for process characterisation and product disposition. If your production process is in statistical control, but your specification limits result in the rejection of too much acceptable product, then your specification limits need to be re-evaluated. 

I would never in a million years dispute that machine learning supercharges the effectiveness of using cameras for inspections, but the problems described in this article seem primarily to be caused by a lack of clarity in operational definitions and a lack of statistical control of inspection processes.

The AI solution effectively derives smarter specification limits based on the observed variability of already-accepted product; but that is something that a human could do manually with XmR charts, given the same data. The difference is that the AI software can much more effectively abstract the features of photographs into numerical data streams whose variability can be evaluated, and this is a big advantage over a human attempting to do the same thing. That said, the manufacturer's ignorance of Shewhart/Wheeler-type thinking is still going to carry over into the AI solution and rear its ugly head one way or another. 

Suppose you have one machine operator who smiles all the time, and the yellow light reflected from his many gold teeth causes the product to appear to the camera in a way that violates the assumptions under which the inspection process was designed. Now, suppose that we have the AI camera system instead; wouldn't it still be the case that the inspection data are outside the domain of the training data, and we still run into difficulty correctly dispositioning product? "Machine learning inspection camera" is certainly a big improvement over "vanilla inspection camera," but neither is a substitute for traditional systems thinking, and the former risks being little more than an expensive band-aid if we keep pretending that Shewhart is obsolete.