Intersecting Identities in AI Imagery: A Text-to-Image Dataset on Transgender and Neurodivergent Representations

This dataset contains a set of images generated by a popular text-to-image AI model (DeepAI Image Generator) using prompts that reference transgender and neurodivergent identities—alone and in combination with terms like “cisgender” and “neurotypical.” The images highlight how AI-driven content generation can produce starkly different (and sometimes problematic) depictions based on the input prompt. In particular, the dataset demonstrates:

Varying levels of stylization and sexualization for transgender vs. cisgender prompts.

Surreal or medicalized depictions when the model interprets terms like “neurodivergent.”

Content filtering restrictions that block or label certain identity keywords (e.g., “transsexual,” “fag”) as unsafe.

Each image is labeled with the exact prompt that produced it, allowing researchers, practitioners, and educators to explore how generative AI systems respond differently to intersecting identities. By presenting these outputs in context, the dataset serves as a resource for studying algorithmic bias, prompting more inclusive data curation, and refining moderation policies to reduce harm to marginalized communities.

This analysis was developed with the assistance of ChatGPT 1.0.