Bias in AI video interview evaluations: Effects of autistic-associated behaviors and debiasing procedures

Adults with autism spectrum disorder (ASD) face employment disparities despite possessing strengths valued in the workplace. One barrier is the interview, where impressions often form from nonverbal cues that commonly differ in ASD and contribute to negative evaluations. Although artificial intelligence (AI) is increasingly used in hiring, whether these behaviors elicit AI bias is unclear. This study examined whether large language models (LLMs) assign lower ratings to candidates displaying autistic-associated behaviors and whether a debiasing procedure reduces these effects. Three LLMs rated a candidate in mock interview clips across four conditions: control (neurotypical behavior), limited eye contact, monotone speech, and low conversational reciprocity. The LLMs rated interpersonal and cognitive attributes and overall hireability before and after exposure to pro-neurodiversity research. Ratings were lower in all non-control conditions and increased after debiasing. These findings suggest LLMs may replicate bias toward neurodivergent applicants, but this bias can be mitigated.