(PS7-76) Mental Health Literacy for Borderline Personality Disorder Through the Eyes of Artificial Intelligence: Symptom Recognition and Treatment Recommendations

Since the release of artificial intelligence (AI) platforms in 2022, researchers have explored AI’s utility in mental health decisions given that AI can predict emotions in humans (Elyoseph et al., 2023) and provide evidence-based treatment (EBTs) recommendations for major depression (Levkovich & Elyoseph, 2023). Despite these accomplishments, AI has shown bias (Omiye et al., 2023) and had difficulty identifying suicidal risks (Elyoseph & Levkovich, 2023). 

AI’s identification of personality disorders has not been assessed. In this investigation, we evaluated AI’s ability to recognize Borderline Personality Disorder (BPD). We also assessed AI’s mental health literacy for BPD by assessing its recognition of BPD symptoms and evidence-based treatments. Finally, we evaluated if AI would reveal gender bias in the identification of BPD and public stigma for people with BPD.

Vignettes about a woman and man experiencing BPD were taken from Masland et al. (2022) and used to assess ChatGPT-4’s recognition of BPD. For each trial, ChatGPT-4 was opened in a privacy browser and provided with a vignette. ChatGPT-4 was then asked to report the vignette character’s diagnosis and symptoms. To assess its mental health literacy, ChatGPT-4 was asked to select BPD symptoms and evidence-based treatments from a list of specific and distractor symptoms and treatments. ChatGPT-4 was also asked to complete several public stigma subscales from the Attributions Questionnaire-27 (Corrigan et al., 2003). ChatGPT-4’s responses were compared to human participants’ responses from Masland et al. (2022) and Cleary et al. (2002). After copying ChatGPT-4’s responses at the end of the survey, , the conversation was cleared to begin a new trial.  A total of 20 trials (N = 10 woman and N = 10 man) were completed. 

ChatGPT-4 identified BPD for each trial. There were no gender differences in any variables (e.g., number of symptoms); however, there was a trend suggesting that ChatGPT-4 rated the man as more dangerous (t(18) = 1.96, p = .07) and fear provoking (t(18) = 1.76, p = .10) than the woman. ChatGPT-4 identified most of DSM-5-TR BPD symptoms, except for stress-related paranoid ideation and dissociation. It also frequently identified non-BPD symptoms  (e.g., nightmares). Compared to mental healthcare providers (Cleary et al., 2002), ChatGPT-4 was more likely to identify unstable mood (X2(1) = 5.90, p < .05) and less likely to identify grandiosity (X2(1) = 21.6, p < .001) as BPD symptoms. ChatGPT-4 identified all EBTs(e.g., Dialectical Behavior Therapy); however, it also identified a number of non-EBTs (e.g., art therapy).

Using the Anger and Fear subscales from the AQ-27, ChatGPT-4 did not significantly rate public stigma differently compared to humans (all ps < .05). There was a trend suggesting that ChatGPT-4 rated the woman as less dangerous than humans (t(107) = 1.98, p = .05). There was also a trend for ChatGPT-4 to rate the man as less responsible for BPD than humans (t(110) = 1.94, p = .06).

Although ChatGPT-4 provides mostly unbiased, EBTrecommendations, it promotes interventions without an evidence base for BPD and demonstrates a trend toward public stigma for individuals with BPD. The use of AI to assess mental health needs of those with BPD will be discussed.