That suggests a level of effort from multiple parties well in excess of the typical "look at item; look at receipt"
Given that the exact timestamp of when the system thinks an item is picked up is known, it should be trivial to review a 5-second clip and flag it as correct or not, training the ML model at the same time.