AI depictions of psychiatric diagnoses: a preliminary study of generative image outputs in Midjourney V.6 and DALL-E...

This paper investigates how state-of-the-art generative artificial intelligence (AI) image models represent common psychiatric diagnoses. We offer key lessons derived from these representations to inform clinicians, researchers, generative AI companies, policymakers and the public about the potential impacts of AI-generated imagery on mental health discourse.

We prompted two generative AI image models, Midjourney V.6 and DALL-E 3 with isolated diagnostic terms for common mental health conditions. The resulting images were compiled and presented as examples of current AI behaviour when interpreting psychiatric terminology.

The AI models generated image outputs for most psychiatric diagnosis prompts. These images frequently reflected cultural stereotypes and historical visual tropes including gender biases and stigmatising portrayals of certain mental health conditions.

These findings illustrate three key points. First, generative AI models reflect cultural perceptions of mental disorders rather than evidence-based clinical ones. Second, AI image outputs resurface historical biases and visual archetypes. Third, the dynamic nature of these models necessitates ongoing monitoring and proactive engagement to manage evolving biases. Addressing these challenges requires a collaborative effort among clinicians, AI developers and policymakers to ensure the responsible use of these technologies in mental health contexts.

As these technologies become increasingly accessible, it is crucial for mental health professionals to understand AI's capabilities, limitations and potential impacts. Future research should focus on quantifying these biases, assessing their effects on public perception and developing strategies to mitigate potential harm while leveraging the insights these models provide into collective understandings of mental illness.