AI Grading in Accounting Reveals Significant Discrepancies
A study highlights systematic differences between AI and human grading in professional accounting exams, noting performance levels and subjects affect outcomes.
An observational study examining the grading of professional accounting exams has revealed interesting patterns of discrepancy between AI and human assessments. The research focused on a task-based machine-learning grading system, which produced results that varied depending on the performance levels of candidates and the subjects examined.
Study Overview
Conducted using archival records from South Africa’s Assessment of Professional Competence (APC), this investigation analyzed 15,507 grading records. These records included grades across eight accounting subjects, namely accounting, taxation, financial management, audit, strategy, decision-making, technology, and ethics.
Methodology
The study involved comparing human grades on long-essay responses with those generated by a supervised machine-learning pipeline specifically crafted for the APC assessment. It’s important to note that this was not a generative large language model. Instead, the system employed a method to convert processed responses into semantic vector representations, ultimately using logistic regression classifiers trained on previously assessed scripts.
Grades assigned by the AI ranged from 2 to 4, with confidence scores provided on a scale of 0 to 100 for each prediction made. Instances where human assessors awarded extreme grades (0, 1, or 5) were flagged for further examination, as were responses deemed insufficient for reliable grading.
Key Findings
The study identified a significant mismatch between AI and human grades when the difference exceeded one point. The analysis utilized various statistical methods, including Welch’s independent samples t-test for gender comparisons and one-way ANOVA for performance levels across subjects.
Notably, the AI's grading showed a stronger discrepancy amongst performance quartiles. In the lower performance quartile, the average AI grade exceeded the human assessment by 1.14 points, while in the higher quartile, the average was lower than human grades by −0.638 points. This demonstrates that lower-performing candidates received higher AI grades compared to their human counterparts.
Discrepancies Across Subjects
Furthermore, discrepancies varied significantly among the eight subjects, indicating that grading consistency was not uniform across disciplines. While factors such as the structure of responses and subject-specific reasoning could potentially explain these variations, the study did not pinpoint exact causes.
Gender comparison did not show a significant impact on grading discrepancies, affirming that AI performance was unbiased with regard to gender. Both male and female candidates received similar treatment, with statistical analysis revealing a lack of significant difference.
Influences on Mismatches
Performance level, subject category, human grading, and AI confidence were all linked to the likelihood of disagreements. Notably, higher confidence levels in AI corresponded to a lower probability of mismatch, highlighting the importance of confidence calibration in grading accuracy.
The study used sophisticated statistical models to emphasize how confidence influenced grading outcomes differently across subjects. The findings showed that confidence values did not consistently predict grading agreement, indicating subject-specific nuances.
Limitations
While the study provides insightful data, it is bounded by certain limitations. It investigated only one machine-learning grading system in a singular context – the South African accounting field. Therefore, the findings cannot be generalized to AI grading systems in other environments.
Additionally, the use of human grades as a predictor may have interlinked statistical dependencies between predictors and outcomes. The restricted AI grading scale of 2 to 4 could exacerbate observed discrepancies, making it difficult to ensure comprehensive calibration.
In conclusion, the study sheds light on the complex dynamics between AI grading and human assessment, emphasizing the importance of ongoing research to understand these discrepancies better and improve grading consistency in the professional accounting domain.
