A new AI tool claims to "personalize risk scores," but its predictions for minority groups are consistently inaccurate. What problem does this illustrate?