Explanations and Machine Learning: Are We Learning Anything?

Thomas M. Menino Convention & Exhibition Center 

The growing field of interpretable machine learning promises to open the black box of modern predictive algorithms, offering human-understandable explanations for complex models. Yet the meaning, validity, and social consequences of these "explanations" remain deeply uncertain. This talk surveys major approaches to interpretability—from local feature attributions and counterfactual reasoning to surrogate modeling and causal abstraction—and asks what, if anything, they tell us about how models actually behave. We will examine how explanations interact with bias and uncertainty, where uncertainty arises, and how adversarial examples and prompts exploit these weaknesses.

Beyond their statistical properties, explanations that are provided to the subjects of automated decisions also affect subject behavior, inducing distribution shift with ramifications for both model performance and explanation validity.

Machine Learning

Explainable AI

Interpretable Models

Distribution Shift