Artificial intelligence systems use code to optimize efficiency, speed, and accuracy. However, human decisions are shaped by values that emerge from social, cultural, and lived experiences.
A study conducted in collaboration with San Diego State University is investigating whether AI systems can learn human values by observing how people make decisions, rather than relying on a single universal rule programmed into them.
Most AI systems today are trained using reinforcement learning. This is a way to teach machines to prioritize by rewarding them with specific outcomes associated with predefined goals. Although this approach is effective for narrow tasks, it often assumes that the values are static and universal. As a result, AI systems may struggle to reflect the diverse ways in which people consider trade-offs, cooperate, and act altruistically across different cultural backgrounds.
Led by the University of Washington (UW) Rodolfo Cortes BarraganAn assistant professor of psychology at SDSU Imperial Valley tested an alternative training approach known as inverse reinforcement learning. Instead of prescribing what an AI system should value, inverse reinforcement learning allows the AI to infer the underlying values by observing human behavior. Their study was published in the journal PLOS One in December.
To examine whether this method could capture culturally shaped values, the researchers focused on altruism. Participants from two self-identified cultural groups, Latino and White, participated in an online cooperative game in which they had to decide in real time whether to help another player, sometimes at personal cost. The researchers found measurable differences in how altruism was expressed between the two groups, with participants in the Latino group showing higher levels of altruistic behavior in the game.
Screenshot of the Overcook online game used to test altruistic behavior. Players must serve as much onion soup as possible with the option of helping other players who have the disadvantage of having to travel far.
The researchers then trained the AI agent using data from different cultural groups of humans and found that the agent learned distinct patterns that reflected group-level differences in humans.
Then, when the AI system was tested in a new scenario involving charitable donations, the AI applied what it learned from the original task. This ability to generalize learned values beyond a single setting suggests that AI systems can learn decision-making tendencies shaped by culture and carry them into the next context.
flexible approach
The findings provide a path forward for developing more adaptive, human-centered AI systems. Rather than classifying individuals or reinforcing stereotypes, the researchers emphasize that culturally attuned AI should be designed to be context-aware, remain flexible, and operate within strong ethical safeguards.
“At the very least, AI development needs to be sensitive to cultural sources and differences in human values,” Cortés Barragan said. “Paying attention to cultural information is key to reducing friction and achieving more fluid interactions, which can lead to more positive outcomes for people using these systems.”
As AI technologies become increasingly integrated into our daily lives, it is essential to understand how values vary across cultural environments. Systems that fail to account for these differences risk misunderstanding human needs or causing unintended harm.
“It is essential that we examine the potential risks of AI and ensure that this technology is firmly rooted in human values so that people can have a safer and more positive experience,” said Cortés Barragan.
By showing that AI systems can learn culturally influenced values through observation rather than explicit instruction, this research contributes to ongoing efforts to design technologies that better meet human needs. While this study does not claim that AI can perfectly replicate human moral reasoning, it does provide a proof of concept for how it can learn values indirectly through behavior.
Additional research is needed to explore how AI systems can responsibly learn a broader range of values and operate in more complex social environments, the researchers said. As AI becomes more integrated into healthcare, education, and public services, understanding how values change across contexts will be critical to building technology that serves people with dignity and care.
The researchers said:
Ngini Oliveira, a research engineer at the University of California at the Allen School, and Jasmine Li, a software engineer at Microsoft who completed the study as a student at the University of California, were co-lead authors. Other co-authors include Kusha Karvati, a scientist at the Allen Institute who completed the study as a doctoral student at the University of Wisconsin. Rodolfo Cortés Barragan, an assistant professor at San Diego State University, completed the study as a postdoctoral fellow at the University of California. Professor Rajesh Rao of the UC Allen School, Professor Andrew Meltzoff, co-director of the UC Institute for Learning and Brain Sciences, and Katharina Reinecke, professor of the Allen School and director of the UC Center for Globally Informative AI.
