I worked for Amazon in Seattle when Amazon Go first launched. As you'd expect, lots of SDE teams made games out of trying to fool the thing in various ways.

A few attempts were successful early on (passing items back and forth, one person moving something to the wrong shelf and another person picking it up, people dressing in identical outfits, etc), but the success rate in fooling the system was very very low. No method of trying to trick it that I ever heard of worked consistently, and it definitely seemed to get harder to fool the longer the store was open.

At the time I thought that whatever algos were being run on the camera feeds were getting better. Knowing that it was basically all manual, I'm not sure what the explanation is for the store seemingly getting harder to fool over time. Possibly just placebo, or people lost interest in trying so hard to fool it.