Reliability of Audit Evidence in Artificial Intelligence Environments: Towards an Explainability and Traceability Evaluation Framework in Light of the 2026 Proposed Revisions to ISA 500

Abstract

Audit firms increasingly use machine-learning tools to analyze entire populations of transactions, yet their outputs are not in themselves audit evidence. This conceptual study examines how the reliability of AI-generated or AI-assisted information intended to be used as audit evidence can be evaluated systematically in light of the exposure draft of ISA 500 (Revised) issued by the IAASB in August 2026. The draft proposes that auditors evaluate the relevance and reliability of such information but leaves criteria for technology-produced information to judgment. The study combines a targeted critical review of auditing and explainable-AI research, a paragraph-level comparison of extant ISA 500 with the 2026 proposal, and design-oriented framework development. Its contribution is the Explainability–Traceability Audit Evidence Evaluation Framework (ET-AEEF), which organizes the evaluation around five linked dimensions: information provenance and integrity, algorithmic explainability, process traceability and reproducibility, evidential relevance and reliability, and professional judgment with human verification. It maps the proposal's reliability attributes onto observable criteria and applies a gate-and-compensation decision logic calibrated to the evidential role of the output. A hypothetical journal-entry anomaly-detection scenario shows that explainability alone cannot justify corroborative or primary reliance without traceability, and that an explainable, reproducible anomaly score directs further procedures rather than proving misstatement. The framework gives auditors and standard setters a transparent basis for documenting reliance decisions; it has not been empirically validated.