It is true that learning algorithms can easily learn conscious and unconscious bias from their training set. But in this case there doesn't have to be a fancy algorithm involved. I think it is plausible that the analysis here is a simple four-liner of code that would hold up to any scrutiny. This is impossible to verify without access to the data, though, and could be racist hatred that is not based on facts at all. In any case, I believe the assumption that their analysis is correct to be valid enough to base a discussion on it.
It is a real social problem that such messages are self-reinforcing. A hiring manager might take this message from the display and not hire blacks. As a result, these are locked in poverty with a higher likelihood to steal, which in turn would lead to a continuation of these store policies. The cycle is hard to break, especially when a store is able to correctly point out that it would lose money by abandoning the policy.
I would say that the store policy is somewhat reasonable, but they cannot be too obvious about it. In the extreme case you would have the isle labeled "beauty products for blacks" entirely locked. In my opinion that would send messages that are harmful to society too loud and clear. But I don't think that it is necessary to lock up all the beauty products, either. Surely there are some other often-stolen beauty products that could be put in there, so that the display doesn't send and obvious message. After all, there are very few people that will see a locked display and make a statistical analysis, whether products marketed to people of African heritage are overrepresented, as long as you cannot tell at first glance.
