Causality Proposed as Solution to Trustworthy AI Trade-offs

A new position paper argues that causality is essential for resolving conflicts among trustworthy AI objectives like fairness, robustness, privacy, and explainability. The authors reinterpret these trade-offs as incompatible invariance requirements under changes to the data-generating process. They propose that causality offers a unifying framework to understand and soften these conflicts through selective invariance. The paper applies this perspective to both classical ML models and foundation models, aiming to balance performance with multiple trustworthiness goals.