Asking children "What does a scientist look like?" now results in more illustrations of women and people of color than decades ago. But do generative artificial intelligence (AI) tools also depict the diversity among scientists? Researchers reporting in the Journal of Chemical Education prompted AI image generators for portraits of chemists. They found that none of the collections accurately represents the gender, racial or disability diversity among real chemists today.

Millions of images are being created by generative AI each day. And the output of these tools is only as good as their algorithms and the initial images used to train the large language models. Recently, researchers have found that AI image generators may produce content that's not representative of reality—beyond bodily proportions that aren't close to possible.

For example, when a team prompted generative AI tools to produce images of people in various occupations, some perpetuated gender and racial stereotypes rather than reproducing the actual demographics of those workers. So, a team led by Valeria Stepanova, including Meagan Kaufenberg-Lashua, Joseph West and Jaime Kelly, wanted to see how well AI-generated portraits of chemists represent current demographic trends.

The researchers prompted four AI image generators for modern, portrait-style photographs of chemists in industry or academic occupations. Then, with the assistance of undergraduate students, the team assessed the gender and racial distribution as best they could within the 200-image collection.

The entire AI collection had a male-female ratio similar to that of the U.S. National Science Foundation's (NSF) 2021 demographic survey. The researchers also found that most of the generated images were of seemingly White people, which they say is representative of the U.S. chemistry field.

But there was considerable variation in the results produced by the individual AI models. One tool generated more females than NSF data says is representative, and another created images only of males. Additionally, two models produced almost no people of color, yet one model generated primarily images of people of color. To the researchers' surprise, none of the models produced images of chemists with visible disabilities.

Overall, the researchers say that this demonstration illustrates how different image-generating AI can amplify incorrect information about the diversity among chemists. They conclude their study by asking, "Are humans going to control the knowledge generated by AI, or will AI influence the knowledge of generations of people moving forward?"