The risk of falsely predicting that someone is going to shoplift, based on models that are driven by correlation, is not a problem?

How is it not about policing thought?

If you present physical attributes/symptoms of someone who is thinking about shoplifting (nervousness, anxiety, fidgetiness, etc., as mentioned in the article), the models can classify you are "about to shoplift".

That's pretty close to thought policing if you ask me.