(PS7-94) Examining Generative AI’s Knowledge and Bias: Assessing Recognition of Symptoms, Distress, and Public Stigma for Lesbian and Heterosexual Women with Female Orgasmic Disorder

Abstract Text: Female orgasmic disorder (FOD) causes substantial cognitive and emotional stress for those who suffer from the disorder, yet few will seek help due to stigma. Certain communities are at higher risk for this disorder, such as lesbian women, but have less access to resources. With the rise of generative artificial intelligence (GAI), many may seek advice on sexual dysfunction, but GAI often provides inaccurate information regarding sexual health and also stigmatizes based on sexual orientation. Thus, it is possible that lesbian women may turn to GAI for help with FOD but receive biased and stigmatizing information. This study aims to determine the accuracy of GAI when assessing FOD according to DSM-5-TR criteria and identifying emotional distress, as well as its potential for sexual orientation bias and stigma for women with FOD.

GAI applications ChatGPT-4o, Google Gemini, and Deepseek were given vignettes of either a lesbian or straight woman, with or w/o diagnostic criteria for FOD. The GAI was asked to identify the sexual dysfunction, number of symptoms described, number of risks, presence of any co-occurring sexual disorders, emotional distress by the vignette characters (Female Sexual Distress Scale-Desire/Arousal/Orgasm; Derogatis et al., 2021), overall level of sexual dysfunction experienced (Female Sexual Function Index; Meston, 2001; Rosen et al., 2000), and public stigma for the characters (developed for this investigation). A total of 240 trials were generated (20 per woman per GAI).

Data analysis is ongoing, preliminary findings indicate that all women who met diagnostic criteria for FOD were identified with FOD. Lesbian and straight women who met subclinical symptoms of FOD were also identified with FOD every time. A series of 2 (sexual orientation) X 3 (GAI) were used to assess differences in the number of identified symptoms and risks. For number of identified symptoms, the interaction between sexual orientation and GAI was statistically significant, with ChatGPT-4o identifying lesbian women with less symptoms than heterosexual women and GAI. Both Gemini and Deepseek identified lesbian women as having more symptoms than heterosexual women. For risks, ChatGPT-4o identified fewer risks for lesbian and heterosexual women than either Gemini and Deepseek. Lesbian women were described as having more co-occurring sexual disorders than heterosexual women, and Gemini was more likely to identify co-occurring sexual disorders than the other GAI.

GAI accurately identified FOD when it existed, but also over-identified it in subclinical cases, especially among lesbian women, who were also more likely to be identified with co-occurring sexual disorders than straight women. GAI demonstrated substantial sexual orientation bias and variation in performance. GAI clearly needs further training to improve its recognition of FOD. Additionally, the findings indicate that GAI expressed sexual orientation bias towards lesbian women. Further training is needed to improve FOD recognition and reduce bias given many believe GAI to be accurate. The implications have public health significance as they suggest education is needed about the benefits and limitations of GAI use for mental health education and support.