The current public company audit standard setting process in the United States is slow and prone to reactiveness, rigidity, and episodic consultation. Drawing on principles from the software industry's agile development process, we discuss how the standard setting process could be improved by breaking large standard setting projects into modular "sprints," forming cross-functional teams, using pilot sandboxes, and embedding continuous investor, auditor, and public feedback. We discuss how governance safeguards, such as quality checkpoints, transparent documentation, balanced advisory panels, and post-implementation reviews, can address risks of rushed or captured rulemaking. By treating standards as living products rather than episodic pronouncements, audit regulators can better protect investors amid rapid technological and market change.
We provide large-sample evidence on the prevalence, determinants, and capital market consequences of artificial intelligence (AI) disclosures in U.S. firms' Form 10-K filings from 2017 to 2024. Using comprehensive AI-related keywords and a large language model, we document a gradual rise in AI disclosure prior to ChatGPT and sharp increases thereafter, particularly in Risk Factors. Post-ChatGPT, AI risk disclosures become longer, more specific, and less boilerplate. Firms' disclosure of AI-related information is strongly associated with ex ante exposure to AI technologies and ex post AI-related negative incidents. Examining market reactions, we find no association between these disclosures and information environment quality prior to ChatGPT's release. After ChatGPT, AI disclosure among high-exposure firms is associated with lower post-filing volatility but higher analyst forecast dispersion and error, suggesting divergent interpretations by investors and analysts. Our findings inform ongoing debates on AI transparency and disclosure regulation.
Machine learning (ML) is increasingly used to detect fraud, predict distress, analyze disclosures, and support audit work. Prior reviews map applications or methods but rarely connect accounting problems to ML design, validation, and governance choices. We reviewed 273 peer-reviewed Web of Science Core Collection articles published from 1992 to 2025, combining bibliometric science mapping with an antecedents-decisions-outcomes (ADO) synthesis enriched by a theory-context-method (TCM) lens. The literature clusters around distress and bankruptcy prediction, reporting quality, and fraud, bridged by audit analytics and textual analysis. ML adoption is driven mainly by assurance pressure, unstructured information, and digital process data; however, many studies rely on accuracy-oriented metrics, nontemporal validation, and explainability that is reported rather than tested. We show how accounting conditions should discipline ML evaluation and propose a research agenda for cost-aware, temporally robust, and governable ML in accounting.
ABSTRACT Continuous auditing has been a subject of academic interest for over three decades. Although extensive research has highlighted its potential to enhance efficiency and effectiveness, practical adoption remains limited. This paper explores this discrepancy by engaging with internal audit practitioners from a national Institute of Internal Auditors chapter to capture a real-world perspective on continuous auditing. Using a comparative analysis format based on a structured literature search, we contrast this practical view with prevailing academic perspectives on continuous auditing, offering a representative comparison between professional practice and academia. We identify key areas of convergence and divergence between theory and practice and propose research questions aimed at addressing the gaps revealed. The paper emphasizes the urgent need for stronger collaboration between academia and practice to address the proposed questions and bridge these gaps, promoting research that yields actionable outcomes. Data Availability: The underlying practitioner guideline is publicly available. JEL Classifications: M42; M40; M48.
Voluntary assurance represents a costly signal of financial reporting quality in equity crowdfunding markets, yet investors often fail to distinguish audit from review reports. Drawing on signaling theory and construallevel theory, I argue that this reflects a fundamental cognitive processing problem: financial statement users default to a high-level construal mindset, focusing on abstract conclusions over engagement-level details. I posit that platform-embedded priming instructions prompting users to consider how the report's conclusion was reached can solve this problem by inducing a low-level construal mindset, redirecting attention toward concrete details that carry the signal. In an experiment with 121 participants, only when primed did investors with an audit report show greater reliance and investment willingness than those with a review report. Perceived credibility also mediated this effect. These findings suggest that assurance signal effectiveness depends not only on disclosure content but on how information is cognitively framed through platform design.
