Large Language Models In Auditing Assessments: Comparisons Between Human Learners, Model Scale, and Enhancement Techniques

This study examines whether large language models (LLMs) can perform audit work by comparing their performance with human learners, across model scale and under two enhancement techniques. Drawing on the audit-judgment distinction between declarative and procedural knowledge, the study evaluates seven publicly available LLMs on audit work in two forms of audit assessment. In the knowledge-based assessment, the LLMs answer four undergraduate auditing examinations previously sat by 352 accounting students, as a proxy of human learners. In the data-analytic case assessment, they plan and execute SQL-based substantive procedures to test the occurrence of revenue and the completeness of expenditures. On the knowledge-based assessment, LLMs outperform human learners where large LLMs typically fall between the 78th and 98th student percentiles, and large LLMs outperform small LLMs. On the data-analytic case assessment, no LLM scores above 50% of the rubric maximum, and large and small LLMs do not differ significantly. Meanwhile, two enhancement techniques: self-reflection and agentic workflow do not improve performance on the data-analytic case assessment. These results reveal a knowing–doing gap: LLMs display strong declarative audit knowledge but weak procedural capability, and self-correction without external feedback does not repair knowledge errors. This study discusses implications for audit practice, assessment design and standard setters.