Title: Resolving Invariance Conflicts in Reliable AI: A Causal Approach

Abstract:

As artificial intelligence (AI)—encompassing both machine learning (ML) models and foundation models (FMs)—is increasingly integrated into high-stakes environments, guaranteeing their trustworthiness has emerged as a primary hurdle. A significant complication is that the fundamental goals of trustworthy AI, including fairness, robustness, privacy, and explainability, are difficult to attain concurrently without compromising utility. This position paper contends that causal reasoning is essential for comprehending and managing the trade-offs between performance and the various objectives of trustworthy AI. We support this stance by reframing these trust-related trade-offs as conflicting invariance demands arising from different alterations to the data-generating process. Through case studies drawn from existing literature and a controlled synthetic-data simulation, we demonstrate that causality offers a cohesive framework for explaining the origins of these trade-offs and for mitigating or resolving them via selective invariance. This theoretical perspective is applicable to both traditional ML systems and large-scale FMs. The paper concludes by identifying key open challenges and future opportunities for leveraging causal methods to develop AI systems that are both highly performant and trustworthy.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC