AI4Society Seminar: Angelina Wang

"More Discrimination for Fairness"

As machine learning has proliferated, so too have concerns about fairness and bias. Much of this work defines fairness as the removal of discrimination — an approach that fits neatly within existing methods like constrained optimization and pursuing clearly defined metrics. However, I argue that this framing has gone too far, failing to recognize the realities of our inequitable world, and theorizing in a social vacuum. To build AI systems that work well for everyone, we must discriminate more. In this talk I will share research on three such contexts where we should treat social groups differently: when reasoning, when simulating human participants, and when personalizing chatbot responses. Finally, I will briefly touch on how even forms of oppression themselves should be distinguished, e.g., that racism and sexism are not interchangeable.