AI in the classroom: assessing GPTs’ capability in undergraduate finance tests

This study investigates the proficiency of three leading GPTs, ChatGPT, DeepSeek, and Gemini, in answering final-exam questions for an undergraduate course in Financial Markets and Intermediaries, a specialized finance education context characterized by heterogeneous cognitive and assessment demands. Drawing on Bloom’s revised taxonomy and Cognitive Load Theory (CLT), this study examines whether GPT performance varies across assessment formats requiring different levels of cognitive processing and reasoning. The results revealed significant divergences: Gemini achieved the highest scores, demonstrating strong comprehension and reasoning skills, whereas ChatGPT showed moderate proficiency and DeepSeek underperformed. All models performed better on multiple-choice questions than on open-ended questions, suggesting that GPTs currently perform more effectively in tasks associated with lower-order cognitive processes. Computational questions posed the greatest challenges across all models, supporting theoretical expectations from CLT. The findings contribute to literature on generative AI in higher education showing that GPT proficiency is not uniform across cognitive assessment levels and by proposing a framework linking AI performance to educational assessment theories. Highly proficient GPTs can serve as valuable supplementary tools for students and educators; however, limitations in complex reasoning, computational accuracy, and regulatory interpretation require critical use. Findings should be interpreted with caution given the single-course single-institution setting.