To some degree it's a concern for "problematic" statistical outcomes, but it's also anti-bias training, both explicit and implicit, giving individual decision makers hesitancy when dulling out punishments.