Bold claim: AI can learn a culture’s values by simply watching how people behave, just like a curious child learns from those around them. But here’s where it gets controversial: should we let machines copy the moral codes of any culture they observe? The study from the University of Washington suggests a promising answer, showing that AI systems can pick up cultural values by watching human behavior and then apply those values in new situations.
Overview
Researchers explored whether artificial intelligence can absorb cultural values by observing people from different cultural groups playing a cooperative video game. In this version of Overcooked, players strive to cook and deliver onion soup as efficiently as possible. One twist adds a helper who provides onions to the other player who has farther to travel, introducing a moment where altruism is rewarded or sacrificed. The goal was to see if AI could learn a culture’s degree of altruism by watching how people from two cultural groups behaved.
Key idea
AI systems traditionally learn values from training data. Since cultures differ in what they consider right or important, a universal, one-size-fits-all AI value system can miss important nuances. The UW team tested whether AI could learn values the way humans do—by observing people in their cultural context and absorbing those patterns, rather than being explicitly programmed with a single set of rules.
How the experiment was built
The researchers recruited 190 adults who identified as white and 110 who identified as Latino. Each group operated an autonomous AI agent. The agents were trained using inverse reinforcement learning (IRL): instead of being rewarded for reaching a predefined goal, the AI observes human (or another AI) behavior to infer the underlying goals and rewards that shaped that behavior.
Why IRL matters
IRL mirrors a natural learning process: people often infer others’ goals from observed actions, not from explicit instructions alone. This aligns AI closer to how children pick up social norms—by watching how people act rather than being lectured about rules.
Experiment Details
In the core task, players directed their in-game actions to maximize onion soup while one player could receive help from another who was farther away. The human participants were unaware that one of the “helpers” was actually a bot designed to solicit assistance. The crucial observation was whether participants would share onions with the bot and, if so, how much.
Findings
- Across groups, the Latino participants tended to help more than the white participants.
- The AI agents trained on Latino data adopted a higher degree of altruism and donated more onions in the same setup than the agent trained on white data.
- A second test replaced the game with a monetary donation scenario. Again, the Latino-trained AI demonstrated greater willingness to help.
Interpretation
The results indicate that IRL-based learning can capture culture-specific altruistic tendencies and transfer them to novel situations. The researchers suggest that expanding the diversity and volume of culture-specific data could allow AI systems to be fine-tuned to reflect the values of a targeted community before deployment.
Limitations and questions
Further work is needed to test this approach in real-world contexts with broader cultural groups, competing value systems, and more complex challenges. How well do these models handle conflicting values or rapid cultural changes? Can we ensure safety and avoid stereotyping when deploying culturally tuned AI?
Implications
If scalable, this approach could help create culturally aware AI that better respects local norms and civic expectations. That raises important conversations about how much cultural tailoring is appropriate, who decides which values are taught to AI, and how to balance plurality of perspectives with universal safety standards.
Contributors
Nigini Oliveira and Jasmine Li led the study, with additional contributions from Koosha Khalvati, Rodolfo Cortes Barragan, and Katharina Reinecke, among others, under the Paul G. Allen School of Computer Science & Engineering and the Center for Globally Beneficial AI at UW.
For more details
If you’d like to dive deeper, you can read the full study in PLOS ONE. And consider: as AI becomes more culturally attuned, should creators prioritize regional customization or aim for a broader, more universally acceptable value framework? Share your thoughts in the comments.
