The article makes no mention of racism. The article only mentions "fidgeting, restlessness and other potentially suspicious body language".
1) The "signs of shoplifting" are human-designed, and so are heavily biased 2) The "signs of shoplifting" are entirely machine learning-based, meaning that they are overwhelmingly biased towards the people who actually get caught and the people who are of communities more likely to shoplift, either because of economic necessity or because of being more likely to live in areas with higher crime rates for whatever reason.
Both lead to racist results without any explicit racism being required.
Also in the absolute best-case scenario that this system works, all you're achieving is locking up more people stealing out of desperation for the benefit of corporations. Even that terrible situation, though, is a pipe dream compared to the reality of systems like these.
If the discrimination (not racism!) causes a bias towards a certain race, neglecting other races who shoplift, and focusing on innocents, then the accuracy will suffer. These systems are not deemed problematic because they are ineffective, they are problematic, because they do a "good" job.