These systems are interested in probabilities, not in finding the causes of shoplifting. If not correlation, then what else?
The causation gap for these models presents a really terrible scenario.
Without causation, these models are capable of correlating things like race, gender, etc. with outcomes like shoplifting.
Nobody infringes on your rights, if a system pings you as suspicious and so a security guard is alerted and watches you more closely. There are no false positive consequences here: either they find that can of Red Bull on your person without a paid receipt, or they don't. You won't get banned from the store, because the system inadvertently deemed you suspicious.
A whitebox statistical model still can provide the factors that contributed to its prediction, without resorting to causal inference.
I do not think a certain race, gender, age, income causes shoplifting, but they sure are correlated, and effective at finding shoplifters or tax fraudsters. I deem the act of shoplifting more unethical than the act of watching you on a camera feed. This already happens anyway, these systems just manage attention better.