In my PhD, I studied how humans and machines use language to express morality. Morality is a core component of human cognition and is believed to have co-evolved with language in its sophisticated modern form. Like language, morality is dynamic: moral views shift over time and vary across individuals and cultures. But these changes are often more subtle and harder to detect than familiar forms of linguistic change. My work asks how moral dynamics can be identified automatically and at scale, and how language technologies such as LLMs can be designed to adapt to evolving natural of human morality.

1) How do morals vary over human history and across cultures? My research characterizes this variation through developing psychologically inspired computational frameworks. For example, my work introduces the Moral Association Graph (MAG), a cognitive model based on human semantic memory that reflects people’s intuitive moral associations (e.g., smoking->disgusting,unhealthy,addiction). I also find that MAG can be extended to historical time points using graph neural networks and large-scale diachronic corpora that date back hundreds of years. I am also interested in extending this computational framework to model cultural universals and variation, answering questions such as: Why do some cultures moralize practices like smoking or divorce, while other cultures do not, and can we predict such cultural moral variation by studying people’s mental representations of word meanings?

2) How do AI systems “perceive” human morality and its variation? My research builds on discussions on the development of ethical AI. Particularly, my work raises critical questions regarding cultural moral variation and how biases in AI and other computational methodologies prevent us from understanding human morality at global scale. My work in this domain has pioneered the use of global ethical surveys for LLM cultural evaluations, and finds that AI systems exhibit a bias in capturing a more accurate representation of moral norms in Western cultures and wealthy nations, while their representation of non-Western moral standards contains harmful stereotyping. My recent work suggests that this lack of accurate cultural representation in LLMs is deeply intertwined with how AI systems capture and interact with ethics and human ethical standards. Extending this line of work, I am interested in studying how morality grows in AI systems in comparison to children’s moral development over time, and exploring AI moral perception through multimodal input (like speech, vision, and text).