Calibrated Skepticism: A Conceptual Framework of Auditor Reliance on Artificial Intelligence

Abstract

Purpose: Artificial intelligence (AI) is rapidly entering the financial statement audit, yet the empirical literature reports two contradictory reliance failures: auditors sometimes defer uncritically to AI output (automation bias) and sometimes discount reliable AI evidence after observing its errors (algorithm aversion). This paper develops a theory that explains when each failure arises and how professional skepticism can be preserved in AI-enabled audits.

Design/methodology/approach: The paper adopts a conceptual research design combining problematization, an integrative review of auditing, judgement and decision-making, and human–automation research, theory synthesis and model building. Professional skepticism serves as the domain theory, while trust-in-automation and algorithm-aversion research serve as method theories.

Findings: The paper introduces the construct of calibrated skepticism, defined as the alignment of an auditor’s reliance on AI output with that output’s actual reliability for a given assertion, sustained through skeptical judgement and action. A framework specifies AI-system, auditor and task–context antecedents of calibrated skepticism, eight testable propositions and a typology of four reliance stances: calibrated partnership, blind deference, reflexive rejection and ritualistic compliance.

Practical implications: The framework informs standard setters revising audit-evidence requirements, audit firms designing AI governance and review procedures, regulators planning inspections and educators developing AI literacy in accounting curricula.

Originality/value: The paper reframes over- and under-reliance on AI as two symmetrical failures of professional skepticism rather than as separate phenomena, and offers a research agenda for experimental, archival and qualitative audit research.