Model Multiplicity for Responsible AI

Machine learning has experienced a remarkable rise, with highly sophisticated over-parameterized models leading the way. Consequently, these cutting-edge models find application across diverse domains. Their increasing deployment has sparked concerns about their real-world impact, studied under the umbrella of responsible AI. A crucial aspect of building responsible AI models is the idea of model multiplicity. If managed well, model multiplicity gives us the freedom to prioritize several metrics, including those associated with responsible AI, and select the best models to minimize harm. However, the existence of multiplicity also marks the unavoidable presence of arbitrariness in model selection that can impact individual-level decisions, necessitating a broader discussion on the role and expectations of AI decision makers in our society.