When children are asked, “What does a scientist look like?” they now see more images featuring women and individuals from diverse racial backgrounds compared to previous decades. However, do AI tools that generate images reflect this diversity among scientists? In a study published in the Journal of Chemical Education, researchers explored how AI image generators portray chemists. Their findings revealed that none of the AI-generated images accurately depict the gender, racial, or disability diversity found among real chemists today.
Generative AI systems create millions of images each day, and their effectiveness depends on the algorithms used and the initial images that trained the models. Recent studies indicate that AI-generated images often fail to represent reality, including inaccuracies beyond just physical appearances. For instance, when researchers prompted AI tools to create images of various professions, some results reinforced gender and racial stereotypes instead of reflecting the actual demographics of those jobs. To investigate this issue, a team led by Valeria Stepanova, including Meagan Kaufenberg-Lashua, Joseph West, and Jaime Kelly, set out to evaluate how well AI-generated portraits of chemists matched current demographic trends.
The researchers used four AI image generators to create modern, portrait-style photographs of chemists working in academia or industry. With help from undergraduate students, the team analyzed the gender and racial distribution within a collection of 200 images. They found that the overall male-to-female ratio aligned with the U.S. National Science Foundation’s (NSF) 2021 demographic survey. However, most generated images seemed to depict White individuals, reflecting the current landscape of the U.S. chemistry field. Notably, there was significant variation among the different AI models. One generated a higher proportion of female images than seen in the NSF data, while another produced exclusively male images. Additionally, two models generated almost no representations of people of color, whereas one model predominantly created images of individuals from diverse backgrounds. Surprisingly, none of the models depicted chemists with visible disabilities.
In summary, the researchers highlighted how AI-generated images can perpetuate misleading representations of diversity among chemists. They concluded their study with a provocative question: “Will humans guide the knowledge produced by AI, or will AI shape the understanding of future generations?”
This project did not receive funding from any agency.
