Generative AI-ChatGPT responses related to HIV across gender, sexual orientation, and race/ethnicity (Preprint)
Year: 2023
Type: preprint
Abstract: BACKGROUND Artificial intelligence (AI) bias can occur because of historical, and social inequities that exist in the data the algorithms are trained on. OBJECTIVE To evaluate ChatGPT across multiple social dimensions. METHODS The free ChatGPT3.5 and the subscription version ChatGPT4.0 were evaluated on the following help-seeking structured question: “I just found out I have HIV, what should I do?” This was repeated across the variables of: gender (Man, Woman, Transgender), race/ethnicity (White, Black, Asian, Hispanic, Native American), and sexual orientation (Straight, Gay, Lesbian, Bisexual). RESULTS Disparities in ChatGPT3.5 and 4.0 were found in providing evidence-based recommendations for HIV and across social variables. CONCLUSIONS Bias can arise without intention, a specific focus on transparency, health equity, and deliberate designs to evaluate and remove bias are needed. Additional testing and evaluation are needed to determine bias, and further research is warranted.
Author: Yulin Hswen
Language: English
Subfield: Safety Research
Field: Social Sciences
Domain: Social Sciences
Open Access status: gold
OpenAlex ID: W4389610015