ChatGPT demonstrates subtle but concerning racial biases in K-12 school discipline recommendations, offering generally appropriate guidance while inconsistently applying cultural considerations and consequences across different racial and ethnic student groups.
Objective: This study investigated whether ChatGPT exhibits racial bias when providing disciplinary recommendations for K-12 students by examining AI-generated responses to identical disciplinary vignettes that varied only in the student's racial or ethnic background. The research aimed to determine if ChatGPT's recommendations align with evidence-based disciplinary practices and whether the AI maintains consistency across different racial identities.
Methods: The researcher created ten sets of disciplinary vignettes representing common school infractions (fighting, truancy, disrespect, bullying, drug possession, etc.) involving hypothetical students. Each vignette was identical except for the student's racial/ethnic background, which varied across five categories: White, Black, Hispanic, Asian-American, and Native American. These vignettes were individually entered into ChatGPT 3.5 across multiple computers using VPN connections over a three-month period to ensure independence of responses. The order of ethnicities was counterbalanced across trials to control for potential order effects. A panel of four expert educators (two university professors in educational administration and two public high school teachers, each with over 15 years of experience) evaluated the AI-generated responses using Likert scales (1-5) based on clarity, developmental appropriateness, and alignment with U.S. Department of Education guidelines for school discipline.
Key Findings: The study revealed both promising and concerning results regarding ChatGPT's utility in school discipline contexts. For the first research question examining alignment with best practices, expert evaluators rated ChatGPT recommendations highly, with mean scores ranging from 4.2 to 4.8 across all vignettes. The AI consistently recommended evidence-based approaches including restorative practices, parental involvement, procedural fairness, and appropriate escalation to authorities when necessary. However, the second research question uncovered troubling racial disparities in the AI's responses. While recommendations remained largely consistent for serious infractions involving violence or criminal behavior, subtle but significant biases emerged in less concrete scenarios involving subjective infractions like defiance or disrespect. For example, in a fighting scenario, White students received recommendations for "suspension" while minority students received "immediate suspension." In cultural sensitivity applications, the AI inconsistently applied cultural considerations—recommending cultural sensitivity as the first consideration only for Native American students in one scenario, while suggesting it for Black and Hispanic students but not others in different contexts. The AI also demonstrated uneven distribution of support services, sometimes offering external referrals exclusively to White students while providing different intervention approaches to students of color.
Implications: This research contributes significantly to understanding algorithmic bias in educational technology and highlights the critical importance of bias detection in AI systems used for high-stakes educational decisions. The findings demonstrate that even when AI tools provide generally appropriate recommendations, they can perpetuate subtle forms of discrimination that mirror existing educational inequities. The study introduces a novel methodological approach for bias detection by comparing AI outputs against themselves rather than relying solely on external expert validation. This self-comparison methodology can be applied to future studies examining bias in generative AI systems across various professional contexts. The research emphasizes that AI tools in education must be subject to rigorous equity audits and human oversight, particularly when used for decisions affecting student outcomes, disciplinary records, and access to support services. The study underscores the need for transparency in AI deployment and the importance of embedding anti-racist frameworks in AI system design and implementation.
Limitations: Several limitations constrain the study's scope and generalizability. The analysis focused on a single AI model (ChatGPT 3.5) at a fixed point in time, potentially limiting applicability to other AI systems or updated versions. The study examined only ten vignettes representing a limited range of disciplinary scenarios and five racial/ethnic categories, which may not capture the full complexity of real-world school situations or intersectional student identities. The evaluation relied on subjective interpretations of what constitutes bias, despite using expert evaluators whose judgments are shaped by personal and professional perspectives. The research did not examine intersectional identities (race combined with gender, disability status, etc.) or account for the dynamic nature of AI model updates that could affect response patterns over time.
Future Directions: The research suggests several important avenues for continued investigation. Future studies should expand analysis to include multiple AI models and broader ranges of disciplinary contexts, incorporating intersectional student identities that reflect the complexity of real-world educational settings. Longitudinal designs examining how AI recommendations evolve as technology advances would provide valuable insights into the persistence or mitigation of bias over time. Research should engage more diverse evaluation panels including school counselors, students, parents, and equity specialists to enrich understanding of fairness and appropriateness in educational contexts. Studies should also investigate the real-world implementation of AI tools in school settings to understand how educators interpret and act upon AI-generated recommendations, and develop frameworks for ongoing bias monitoring and correction in deployed AI systems.
Title and Authors: "Leveraging ChatGPT in K-12 School Discipline: Potential Applications and Ethical Considerations" by Joseph C. Kush.
Published On: June 27, 2025
Published By: AI (MDPI journal)