ABSTRACT As artificial intelligence (AI) reshapes higher education, accounting instructors face new opportunities, and challenges, in leveraging these tools for assessment and instruction. One persistent issue in business education is the time-intensive process of generating individualized case materials that are both pedagogically sound and scalable. Although AI models like ChatGPT-4o have been explored for general tutoring or summarization, their potential to automate detailed instructional content, especially in cost accounting, remains underexamined. This paper evaluates the capabilities and limitations of ChatGPT-4o Plus in generating custom cost accounting case studies and self-grading templates, focusing on cost-volume-profit analysis using both traditional and activity-based costing. Through a structured prompt engineering and testing process, we provide a practical roadmap for accounting educators seeking to adopt AI for individualized instruction, while also highlighting critical limitations around scalability and Learning Management System (LMS) integration. Data Availability: The prompts, case materials, and instructional artifacts used in this study are included in the manuscript and accompanying supplemental materials.
ABSTRACT Textual analysis generally refers to the use of automated analysis techniques to describe the content of text-based materials. Although the application of textual analysis in accounting has been explored in academic research, there has been little focus on the extent of adoption of textual analysis in accounting practice itself. The primary purpose of this study is to explore the adoption of textual analysis from the practice side of the accounting profession prior to the emergence of generative artificial intelligence. We perform a two-part study, initially interviewing a variety of accounting-related professionals and then testing a theory-based research model associated with technology adoption. Our results suggest that, although the adoption of textual analysis by accounting professionals is driven by the traditional factors of effort expectancy and performance expectancy, an important factor in the adoption and acceptance of textual analysis is the support of professional bodies associated with the accounting profession. Data Availability: Data are available upon request. JEL Classifications: M41.
ABSTRACT Financial fraud threatens capital markets and audit practice. This study examines whether auditors’ narrative comments in Chinese-language annual reports contain affective cues that provide early warning signals of misrepresentation. Using multiple sentiment dictionaries, affective features are extracted, and an integrated machine learning framework is developed that combines sentiment-based clustering with class-imbalance resampling. Five resampling strategies are compared, including random oversampling (ROS), random undersampling (RUS), the synthetic minority over-sampling technique (SMOTE), SMOTE with edited nearest neighbors (SMOTE + ENN), and SMOTE with Tomek links (SMOTE + Tomek), across logistic regression (LR), support vector machine (SVM), random forest (RF), and k-means clustering models. The SMOTE-based resampling substantially improves minority-class detection. The results show that emotional language in auditors’ narratives has predictive value for fraud risk and provide scalable implications for audit analytics and policy across languages and regulatory settings. Data Availability: The annual reports used in this study are publicly available from the Taiwan Market Observation Post System (MOPS), and compensation-claim data are publicly available from the Taiwan Securities and Futures Investors Protection Center (SFIPC). Processed data and code are available from the corresponding author upon reasonable request, subject to data-use restrictions. JEL Classifications: C45; C55; M41; G32.
ABSTRACT This paper presents the design of an academic Lab that leverages Deep Reinforcement Learning (DRL) to bridge research, education, and industry. By applying Lean Six Sigma, the lab structures domain expertise as Markov Decision Processes (MDPs) to enable data-driven AI development. A four-layer architecture supports continuous data collection, adaptive foundation model training, and knowledge transfer. Unlike data-intensive LLMs, the proposed adaptive risk analytics model is built on theoretical domain knowledge and validated with small datasets. Through meta-learning, the lab accumulates and transfers intelligence across tasks and disciplines. Functioning as a knowledge-augmented simulator, it facilitates dynamic DRL training, LLM integration, and practical applications in auditing, accounting, and business. A case study illustrates how the approach accelerates the translation of theory into solutions while strengthening industry-academia collaboration. Data Availability: The dataset used in the case study is available from the author by request.
Channel stuffing has been a focal point for both management and regulators. Managing sales activities across countries and ensuring accurate recording of sales is challenging for management, especially in multinational corporations. This study develops a state-of-the-art channel stuffing detection framework from an internal perspective using statistical analysis alongside unsupervised outlier detection techniques. The primary goal of this framework is to identify unusual sales activities related to channel stuffing and ensure that sales activities are conducted per the company's compliance policy. This framework provides management with a holistic and granular view of sales activities at the country and customer levels, indicates unusual trends and patterns of transactions, and underscores high-risk channel stuffing activities for further investigation. The framework also enhances the company's monitoring procedures and produces tangible applications.
