Bold claim: AI should learn culture the way children do, by watching people live their values, not by memorizing a single universal code. And this is the part most people miss: culture isn’t one-size-fits-all, so the way we teach AI to behave must reflect diverse norms. A new study from the University of Washington explores exactly that idea, showing how AI can acquire cultural values by observing human behavior rather than by hard-coding a single set of rules.
The core challenge is clear: AI systems trained on broad internet data inherit a mix of values, which may not line up with every cultural context. If an AI is built to operate effectively across cultures, it can’t rely on a monolithic value framework. The UW team tested a different approach. They trained AI agents by watching people from two distinct cultural groups play a cooperative video game, with the goal of understanding how group-specific altruism translates into decisions in new, unfamiliar situations.
What they did and what they found
Researchers recruited 300 participants—190 identifying as white and 110 identifying as Latino—and assigned each group an autonomous AI agent. The agents learned through inverse reinforcement learning (IRL). Unlike standard reinforcement learning, where the AI is rewarded for achieving a predefined objective, IRL lets the AI infer the underlying goals and rewards by observing human behavior. This mirrors how humans often learn by watching others and inferring intentions rather than being told explicitly what to do.
To test cultural learning in a concrete setting, the team used a modified version of Overcooked, a cooperative cooking game. In this scenario, players prepare onion soup and must coordinate with a second player who is positioned at a disadvantage. Unbeknownst to participants, the second player was a bot that could request help. Participants faced a personal cost if they chose to share onions with the bot, since helping the bot meant delivering less soup themselves.
Findings showed that the Latino participants, on average, were more likely to help others than the white participants. Importantly, the AI agents trained on each group’s data internalized that group’s altruistic tendencies. In the Latino-trained AI, the agent donated more onions to the bot in need, an action that reflected the observed cultural pattern. A follow-up test confirmed that the Latino-trained agent also demonstrated greater generosity when faced with a separate charitable donation task, indicating a broader, transfer-ready value pattern rather than a behavior limited to the game.
Why this matters
These results suggest that IRL-based training could allow AI systems to adopt culture-specific value profiles by simply observing people within a culture, rather than relying on a universal ethical blueprint. As Rao notes, this approach could enable AI companies to fine-tune their models for particular cultural contexts before deployment, potentially reducing misalignment with local norms. However, the researchers caution that real-world deployment would require more data across additional cultural groups and more complex scenarios to understand how competing value systems interact.
A broader takeaway is that culturally attuned AI is a societal imperative. If AI is to operate effectively and ethically in diverse communities, it must be capable of adopting perspectives beyond a single universal standard. Meltzoff emphasizes that human values are often learned through subtle social absorption—what he describes as learning to act by osmosis within a community rather than through explicit instruction.
Who contributed and where to learn more
Nigini Oliveira, a UW post-doctoral researcher, and Jasmine Li, a Microsoft software engineer who contributed to this work as a UW student, co-led the project. Additional collaborators include researchers from UW, the Allen Institute, San Diego State University, and other institutions. The full study is published in PLOS ONE.
If you’re curious about this line of inquiry, consider this question: Should AI systems be tailored to the moral sensibilities of the communities they serve, or should they adhere to a universally applicable ethical framework? What balance between cultural adaptability and universal safety standards would you find most responsible—and why?
