Identity and the limits of fair assessment

In many assessment problems—aptitude testing, hiring decisions, appraisals of the risk recidivism, evaluation credibility testimonial sources, and so on—the fair treatment different groups individuals is an important goal. But can be legitimately grouped in ways. Using a framework fairness constraints explored research on algorithmic fairness, I show that eliminating certain forms bias across for one way classifying make it impossible to eliminate such another dividing people up. And this point generalizes if we require merely assessments approximately bias-free. Moreover, even are satisfied some given partitions population, fail coarsest common refinement, is, partition generated by taking intersections elements these coarser partitions. This shows prominent admit possibility intersectional bias.