An SLP Without the Degree: Investigating Demographic Disparities in AI-driven SLP Eligibility Decisions of School-Aged Children

Authors: Adriana Gleeson, Sasha Gozan-Keck, Samiha Kazi

Field of Study: Psychology

Program Affiliation: Binghamton University Projects for New Undergraduate Researchers (BUPNUR)

Faculty Mentors: Gregory Hallenbeck, Cassandra Natali

Easel: 92

Timeslot: Morning

Abstract: As AI is increasing its presence in healthcare, it is important to identify its potential dangers such as racial, gender, and social biases. It was hypothesized that large language models (LLMs), such as ChatGPT, reflect these biases when determining the eligibility of school-age children for SLP services. A scenario describing a child displaying various features of African American English (AAE) was given to Chat GPT for children of varying race, sex, and socioeconomic status. It was asked to determine eligibility of SLP services and provide reasoning. After collecting data on 500 variations, a logistic regression identified any significant main effects. It was found that ChatGPT was more likely to recommend services for students of certain races and socioeconomic statuses. ChatGPT also made assumptions about language acquisition and language exposure. Being aware of the potential biases that LLMs produce is important for SLPs to use this tool ethically and responsibly.