The Myth in the Methodology: Towards a Recontextualization of Fairness in Machine Learning - green_icml18.pdf

Even as machine learning has expanded into therealm of social decision-making, where concernsof bias and justice often rise above those of effi-ciency and accuracy, the field has remained com-mitted to standard ML techniques that conceiveof fairness in terms of statistical metrics and relyheavily on historical data as accurate and neutralrepresentations of the world. So long as the fieldconforms to these methods and believes it can op-timize systems according to universal notions offairness, machine learning will be ill-suited to ad-dress the fundamentally political and ethical con-siderations at stake when deploying algorithmsin the public sphere. The design and adoption ofmachine learning tools in high-stakes social con-texts should be as much a matter of democraticdeliberation as of technical analysis.