Aliah Zewail Co-authors New Research on Moral Stereotyping in Large Language Models

Content

Image

Aliah Zewail, a graduate student in the Department of Psychological and Brain Sciences.

Aliah Zewail, a graduate student in the College of Natural Sciences’s Department of Psychological and Brain Sciences (PBS), has co-authored a new paper, “Moral Stereotyping in Large Language Models,” in the Proceedings of the National Academy of Sciences (PNAS) that examines the confluence of artificial intelligence (AI), large language models (LLMs), morality, and cultural diversity. This paper was written in partnership with: Alexandra Figueroa, a postdoctoral scholar at the University of California, Berkeley, Haas School of Business; Jesse Graham, the George S. Eccles Chair of Business Ethics and Professor of Management at the University of Utah; and Mohammad Atari, an assistant professor at PBS. 

“We provide a cautionary tale of researchers using LLMs as generators for cultural data,” Zewail says of the paper. “Specifically, we provide insight into the value structure of LLMs and how they misrepresent the moral profiles of less-WEIRD (Western-educated, industrialized, rich, and democratic) nations, portraying them as less morally concerned than they actually are.” 

The paper goes on to argue that generative pre-trained transformers (GPTs)—such as ChatGPT—play a role in this misrepresentation of the moral values of non-Western individuals. “We showcase that social scientists should not replace human participants with LLMs, as these AI systems fail to capture human diversity around the globe,” Zewail asserts. 

A map from the paper illustrating GPT-4o’s deviations from humans’ actual overall moral concerns. Results show that GPT-4o was more accurate in estimating the overall moral concerns of Western nations.

This paper comes on the heels of Zewail receiving a grant from the National Science Foundation Graduate Research Fellowship Program (NSF-GRFP), which is supporting her research on moral values and the cultural evolution of democratic norms in the Middle East and North Africa. 

“The application of this paper is critical as AI systems increasingly assume more involvement in some researcher (and laypeople) decision-making processes,” Zewail contends. “By addressing potential inaccuracies in a model’s output, their work aims to inform policy that prompts ethical AI development and usage to enhance the reliability, accuracy, and cultural competence of what is becoming a ‘global’ tool.” 

Click here to read the paper in PNAS.