A new study from the University of Washington suggests that artificial intelligence systems may be able to absorb cultural values the same way children do — by simply watching how people behave.
Published on December 9 in PLOS One, the research shows that AI trained on human decision-making can pick up varying degrees of altruism across cultural groups and apply these learned values in entirely new situations.
Why Cultural Values Matter in AI
Most AI systems are trained on massive datasets gathered from across the internet, which means they inherit a mishmash of global values. But cultural norms differ significantly across societies, making a one-size-fits-all value system ineffective or even harmful.
Lead researcher Rajesh Rao, professor in UW’s Paul G. Allen School of Computer Science & Engineering, explains:
“We shouldn’t hard-code a universal set of values into AI systems. Cultures shape values differently. We wanted to see whether AI could learn values in the same way children do — through observation.”
The inspiration came from earlier UW research showing that children from Latino and Asian families displayed higher altruism at 19 months of age than children from other cultural backgrounds.
How AI Learned Altruism Through a Video Game
Researchers asked two groups of adults — 190 white participants and 110 Latino participants — to play a modified version of the cooperative game Overcooked. Players cooked onion soup while seeing that an AI-controlled “partner” in a nearby kitchen was at a disadvantage and occasionally asked for help.
Players could choose to give away onions to help — sacrificing their own score — or keep playing competitively.
Key Findings
- Latino participants helped significantly more often than white participants.
- AI agents trained on each group’s gameplay learned to mirror the same level of altruism.
- In a later, unrelated money-donation test, the agents again displayed the same learned values, proving they had internalized a general principle of altruism, not game-specific behavior.
This was achieved through inverse reinforcement learning (IRL) — a method where AI observes human behavior and infers underlying motivations rather than being explicitly trained to maximize rewards.
Learning Values Like Children
Co-author Andrew Meltzoff, co-director of the UW Institute for Learning & Brain Sciences, explained the human parallel:
“Kids don’t learn values through repeated training. They absorb them by watching parents and community members. Values are more caught than taught.”
The study shows that AI systems can develop similar implicit learning and may be able to adopt culturally aligned behavior if given culturally specific examples.
Why This Matters
The findings could help companies tailor AI assistants, robots, or automated decision systems to the norms of the communities they serve.
Rajesh Rao says:
“With enough behavior data, AI systems could be fine-tuned to reflect the values of a specific culture before deployment.”
But researchers caution that real-world scenarios involve:
- multiple, sometimes conflicting cultural values
- complex ethical dilemmas
- large, diverse populations
Much more work is needed before culturally attuned AI systems can be safely deployed.
The Bigger Question
Co-author Meltzoff emphasizes the long-term importance:
“Creating culturally attuned AI is an essential societal question. How do we design systems that can take others’ perspectives into account and behave in civic-minded ways?”
The study represents an early but promising step toward AI that understands — and respects — cultural differences.
