What Does AI\nThink a Chemist Looks Like? An Analysis\nof Diversity in Generative AI

Four artificial intelligence (AI) image generators (Adobe\nFirefly,\nDALL·E2, Craiyon, and DreamStudio) were utilized to produce “faces”\nof chemists with varying occupational titles. Images were analyzed\nfor representational biases and compared to data available from National\nScience Foundation (NSF). Presentational biases were analyzed within\nAmerican Chemical Society (ACS) Diversity, Equity, Inclusion, and\nRespect guidelines with a particular focus on presentation of diversity\nand disability. Amplification of both representational and presentational\nbiases was observed for all four AI generators despite alignment of\ndemographic trends with NSF and ACS reporting. Influences of occupationally\ntied prompts (specific to chemistry) on demographic distributions\nof AI-generated images were investigated. At least one AI image generator\nassigned women and racial minorities to “assistant”\npositions while males and whites occupied the “top”\npositions in the field. Our data also demonstrates erasure of people\nwith visible disabilities in the AI-generated outputs. Perspectives\nof current students in chemistry classes at Winona State University\non “what a chemist looks like” have been collected and\nanalyzed. The disturbing prevalence of a “white male”\nimage of a chemist, even for students identifying as female or a person\nof color, was evidenced. This research brings forward concerns about\nthe utilization of “black box” algorithms for training\ngenerative AI for nonbiased image generation using consumer feedback\nwhen novices themselves perpetuate both representational and presentational\nbiases prior to their exposure to AI’s perspectives, especially\nin chemistry where substantial NSF and ACS efforts have been forwarded\nto diversify and revitalize the next generation of scientists.