AI learns cultural values ​​by watching humans play video games

Artificial intelligence continues to stumble when it comes to values, not because it is incapable, but because not all people have the same values. A system trained across the internet absorbs all the values, so it won't work equally well for everyone. This raises an uncomfortable question. How can we build AI that respects cultural differences without hard-coding a moral rulebook that no one agrees with?

Researchers at the University of Washington think they've found the answer, and it appears to be a lot like how children learn to become decent people. A series of experiments published in PLOS One on December 9 shows that just by observing human decisions in video games, an AI agent can absorb culture-specific ideas about altruism and apply those values ​​to entirely new situations.

The research team's key finding is that AI can recognize implicit human values ​​from observation alone. As participants in one group displayed more supportive behaviors, the AI ​​assigned to them absorbed that group's values ​​and brought them into scenarios never seen before. There are no explicit instructions. There are no moral lectures. Just watch and learn.

Onion soup experiment

Researchers recruited 300 U.S. adults, of whom 190 identified as white and 110 identified as Latino. These groups were chosen based on previous research suggesting cultural differences in altruism, but the researchers stress that this is just a starting point for understanding how cultural learning works in AI systems.

Participants played a modified version of Overcooked, a frenetic cooking game where players compete to prepare and serve onion soup. The twist was fairness. While one player had an easy setup, the other player had to walk further to collect materials. Players with an advantage can share onions to help struggling players, but doing so wastes valuable time and reduces their own score. Support was voluntary, tangible, and costly.

On average, Latino participants chose to help more often than White participants. The AI ​​agents were each trained on one group of behaviors and learned directly from what they observed. The agent trained on data from the Latino group gave more onions than the other agents, mirroring human patterns.

But here's where it gets interesting. The researchers did not use standard reinforcement learning, which gives the AI ​​a goal and rewards it for progress. Instead, inverse reinforcement learning (IRL) was used. AI monitors human behavior and infers the underlying goals and rewards that drive that behavior. It does not learn what humans have done. Learn why they did it.

“Children learn how people behave in their communities and cultures almost by osmosis. The human values ​​they learn are 'caught' rather than 'taught.'” – Andrew Meltzoff, University of Wisconsin Psychology Professor and Co-Director of the Institute for Learning and Brain Sciences

Beyond the kitchen

To confirm that the AI ​​wasn't just memorizing the game's tactics, the team ran a second experiment. They put the agents in a completely different scenario. You decide whether to donate money to another agent in need, knowing that your resources are limited and your future expenses are unpredictable.

No new training was provided. The agent was applying previously learned values ​​to new situations. Again, the AI ​​trained on data from Latino participants acted more altruistically and donated more frequently than the AI ​​trained on data from white participants.

Frankly, this ability to generalize is important. This suggests that engineers do not need to explicitly define norms and that AI systems can learn cultural norms beyond a single task.

“Many cultures have their own values, so a universal set of values ​​should not be hard-coded into AI systems.” – Rajesh Rao, Professor of Computer Science and Engineering, University of Washington

Rao, the study's senior author, said it was a proof-of-concept demonstration. AI companies could potentially fine-tune their models to learn the values ​​of a particular culture before deploying the system within that culture. However, the research team is clear about the limitations. The experiment included only two cultural groups, one value system, and a simplified environment. Real-world settings include conflicting values, higher stakes, and far messier signals.

There are also risks. AI that learns from human behavior can absorb bias, bias, or harmful norms if they exist in the data. Researchers argue that cultural learning requires a combination of monitoring, transparency, and safeguards. Still, this research reframes the central challenge in AI development. Rather than asking what values ​​machines should follow, we ask how machines learn values ​​in the same way humans do, by observing, participating, and adapting within communities.

PLOS One: 10.1371/journal.pone.0337914