Artificial intelligence systems can learn cultural values by observing human behavior, much like children do, according to a new study from the University of Washington. The research, published in PLOS One, shows that AI agents trained on gameplay data from two distinct cultural groups developed different levels of altruism based on the behaviors they observed. The study used a modified version of the video game Overcooked, where players work together to cook and deliver onion soup. In the setup, one player could see a second player in a different kitchen who had to walk farther to complete the same tasks. The second player was actually an AI bot that would request onions from the human player. Helping the bot meant sacrificing personal progress, making it a test of altruism. Researchers recruited 190 white and 110 Latino adults to play the game. On average, participants from the Latino group were more likely to give away onions, indicating higher levels of altruism. The AI agents trained on data from each group learned to mirror the behavior of their respective cultural group. The agent trained on Latino data gave more onions than the one trained on white data. The AI systems used a method called inverse reinforcement learning (IRL), where the system observes human actions and infers the underlying goals and values. Unlike traditional reinforcement learning, where rewards are explicitly programmed, IRL allows AI to learn by watching how people act in real-world situations—similar to how children absorb values by observing their families and communities. “Parents don’t just drill children on specific tasks. They model behavior—sharing, caring, helping others—and kids learn these values by osmosis,” said co-author Andrew Meltzoff, a UW psychology professor. “Our AI approach mimics that natural learning process.” To test whether the AI had truly internalized a cultural value, researchers ran a second experiment. In a new scenario, the agents had to decide whether to donate part of their earnings to someone in need. Again, the agent trained on Latino data showed greater willingness to help, suggesting it had learned a general principle of altruism, not just a specific behavior. Senior author Rajesh Rao, a professor at the UW Paul G. Allen School of Computer Science & Engineering, emphasized the importance of avoiding one-size-fits-all values in AI. “We shouldn’t hard code universal values into AI,” he said. “Different cultures have different values. Instead, AI can learn them by observing people in context.” The findings suggest that AI systems could be fine-tuned to reflect the values of specific cultural groups before deployment, making them more effective and ethical in diverse societies. However, more research is needed to test this approach across more cultures, with complex social dynamics and conflicting values. “Creating culturally attuned AI is a crucial challenge for society,” Meltzoff said. “We need systems that can understand and respect different perspectives, and act in ways that are not just smart, but also civic-minded.”