Statistical Fairness Analysis of Generative Ai-based Educational Assessment Systems for Students with Disabilities

Abstract

Generative Artificial Intelligence (GenAI) has started to play significant role in the field of Education (EDU), especially in Education Assessment & Feedback, Automated Grading & Education Personalisation. Its increasing popularity, however, has sparked a discussion about the need for fairness in determining students' abilities that is rooted in artificial intelligence for evaluating student performance. It’s growing adoption, however, has raised the concerns of whether or not AI-based assessment would be equitable for disabled students. This article discusses, through a secondary review (literature-based review), the statistical fairness of educational assessment systems that make use of GenAI. It reflects important factors of fairness such as demographic parity, equal opportunity, equal odds and disparate impact, and explores the potential for differences in communication, behavior, response time and assistive technology related to disability to impact AI assessments. Based on the reported literature reviewed, overall accuracy of a strong, all-purpose AI does not always lead to equality of results for disability groups. Valid concerns range from the lack of disability-inclusive data and access issues, to model bias and issues with implementing traditional measures of fairness for different disability groups. The article suggests regular fairness audits, accessibility of the assessment design, inclusion of the necessary accommodation features, enhanced disability representation in assessment datasets and ongoing manual verification to facilitate equitable and reliable assessment using intelligence tools.