AI Bias in Special Education

Despite the rapid deployment of artificial intelligence (AI) to support K–12 educators in improving their practice, serious concerns remain unanswered about large-language model (LLM) output. This is especially true in special education, where educators report little training in disability theory, leading them to consult with AI systems without the background knowledge needed for objective analysis of the chatbot responses. To address this issue, we conducted a qualitative content analysis of the transcripts of responses from five major LLMs (Anthropic Claude, Google Gemini, OpenAI ChatGPT, Perplexity Sonar, and xAI Grok) to common questions from early-career educators. We found an overreliance in the chatbot output on the Medical Model of disability, as well as almost no engagement with the Human Rights Model. We also found a wide variability between the different LLMs, as well as internal variability within each LLM based on the wording of the prompt. We conclude by calling for explicit instruction on disability theory in teacher preservice programs to support critical engagement by teachers with AI tools, as well as improved LLM models that consider multiple models of disability. This article makes a practical contribution to special education practice by identifying imbalanced AI output and suggesting ways to minimize the negative impacts that may result from using these tools.