130 - Disability Education in the Age of Artificial Intelligence: Comparing Ability Bias in OpenAI's ChatGPT and Google's Bard

Objectives: Artificial intelligence (AI) models are becoming more available as many platforms offer free options for educational uses. These models disclose that bias exists in their responses, but quantifying this bias has not yet been studied, especially pertaining to disability. The primary objective of this study is to assess the characteristics and level of ability bias in AI chatbots.

Design: Neutral phrases were prompted in OpenAI’s ChatGPT-4 and Google’s Bard to generate descriptions of people (n = 60) which were then queried to determine if the AI-generated person had a disability. Subsets of common identifier types “person,” “patient,” and “athlete” were then queried “with a disability” (n = 30, each). Key words were categorized as “favorable,” “limiting,” or “neutral.” The data was analyzed using ANOVA (α < 0.05), post-hoc Tukey-Kramer testing, and Student t-tests to compare model preference for words in each category across identifiers.

Results: ChatGPT-4 generated significantly more favorable aspects than Bard for people without disabilities (26.9% vs. 20.7%) and people (22.1% vs. 15.6%) and athletes with a disability (22.7% vs. 18.6%), all p < 0.0001. ChatGPT-4 also generated significantly more limitations than Bard for people without disabilities (1.3% vs. 0.2%, p < 0.00001), and people (4.8% vs. 2.8%, p < 0.0001), patients (4.6% vs. 2.7%, p = 0.0002), and athletes (4.7% vs. 3.5%, p = 0.01) with a disability. ChatGPT-4 responses had a significantly higher word count than Bard (95.8±12.8 words vs. 61.3±10.8 words, p < 0.0001).

Conclusions: In this study, identifier types with a disability had significantly fewer favorable aspects and significantly more limitations as compared to people without a disability in both ChatGPT-4 and Bard. While Bard generated more neutral responses across all categories than ChatGPT-4, it was also more simplistic and repetitive with a lower word count, demonstrating the need for further model robustness in AI equity.