Why Auditability Fails in Current Generative AI Systems: A Governance Perspective

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Regulatory frameworks increasingly assume that generative AI systems are auditable. Yet mainstream generative AI architectures do not structurally support ex post reconstruction of system behavior for enforceable legal or regulatory oversight. This review argues that auditability fails because prevailing deployment models lack the infrastructural preconditions necessary for verifiable accountability. Drawing on literature in algorithmic accountability, AI auditing, foundation model governance, and regulatory compliance, the paper identifies structural barriers including probabilistic non-determinism, continuous model updates, API-based black-box provisioning, distributed responsibility across AI supply chains, and insufficient prompt–response trace binding. The analysis distinguishes transparency, explainability, traceability, and auditability, arguing that enforceable auditability requires immutable and reconstructible system evidence. Using a structured narrative review, the paper conceptualizes auditability as a foundational computational primitive from which accountability, compliance, and regulatory oversight can be systematically derived. This concept is operationalized through the AIGov framework and the implemented AIGov Engine, a governance-native reference architecture based on append-only evidence logging, policy-constrained execution, and deterministic compliance-state derivation. The AIGov Engine further demonstrates that these principles can support runtime governance and oversight of autonomous and multi-agent systems.