Some (most?) of these aren't really AI-based at all. For example. traditional optical sorters typically rely on the reflectivity of materials at one or a few laser wavelengths directed onto the material.

The mapping between sensor signals and material types is usually hardcoded from laboratory test results.

Using AI for image recognition is to visual detection as Orange Juice is to beverages.

Sure. What they will call AI [an unspecified number of years in the future] will be compared similarly the the SOTA AI models.

For long time the term "artificial intelligence" has just gone out of favour, but I do remember the days where a good AI research lab had a bunch of symbolics lisp machines

I mean that we have visual recognition systems that do not use any kind of machine learning whatsoever and those are the majority of systems in use at industrial scale.

Laser interferometry and DCT image distance, primarily.