Intrinsic Barriers to Explaining Deep Foundation Models

Arizona State University, USA Deep Foundation Models (DFMs) offer unprecedented capabili ties but their increasing complexity presents profound challenges to understanding their internal worki ngs - a critical need for ensuring trust, safety, and accountability. As we grapple with explaining these sys tems, a fundamental question emerges: Are the difficulties we face merely temporary hurdles, awaiting more sophisticated analytical techniques, or do they stem from intrinsic barriers deeply rooted in the nature of these large-scale models them selves? This paper delves into this critical question by examining the fundamental characteristics of DFMs and scrutinizing the limitations encountered by current explainability methods when confronted with this inherent challenge. We probe the feasibility of achieving satisfactory explanati ons and consider the implications for how we must approach the verification and governance of these powerful technologies. Introduction Deep Foundation Models (DFMs) - such as large language models a nd multimodal architectures - are a class of neural networks trained on vast amounts of data, de signed to serve as general-purpose engines for downstream tasksacross diverse domains [10].With the emergence ofsystems like GPT, Gemini, and CLIP, artificial intelligence is undergoing aprofound transformation.