Why Probabilistic AI Resists Auditability: From Explainability to Verifiable Responsibility in High-Risk AI Systems

Current approaches to AI governance assume that accountability can be achieved through explainability, documentation, and post-hoc controls. This assumption breaks down for probabilistic systems such as large language models, where outputs are non-deterministic, non-reproducible, and only partially attributable. As a result, responsibility is typically assigned retrospectively, after harm has occurred, rather than being technically enforced during decision generation. This creates structural gaps between regulatory expectations—such as auditability, traceability, and liability—and the operational reality of modern AI systems. This paper argues that accountability in high-risk AI systems cannot reliably emerge from model-centric techniques alone. Instead, it proposes a conceptual shift toward system-level responsibility architectures that treat verification, non-decision, and provenance as first-class properties of the execution pipeline rather than as interpretative overlays. The contribution is intentionally non-implementational. It outlines architectural principles for verifiable responsibility in non-deterministic AI systems without disclosing technical specifications or deployment details. The aim is to inform regulatory, legal, and governance discussions on the structural limits of explainability and the requirements for audit-ready AI infrastructure. Context note:This paper was first released via SSRN as a working paper (Abstract ID: 6103587).The Zenodo version is identical in content and provided to ensure long-term archiving, DOI-based citation, and ORCID integration.The contribution is conceptual and non-implementational, intended to inform regulatory and governance discussions on high-risk AI systems.