New research from the University of Washington asks whether artificial intelligence can learn cultural values by observing human behaviour, similar to how children learn. The findings, published in PLOS One, build on prior UW work that found 19-month-old children from Latino and Asian households showed greater altruism on simple tests.
The study recruited adults who identified as white (190 participants) and adults who identified as Latino (110 participants). The team trained separate agents with inverse reinforcement learning (IRL). In contrast to standard reinforcement learning, where an agent receives an explicit goal and rewards, IRL infers goals and rewards by watching human actions — an approach the authors say more closely matches how parents model behavior for children.
In a modified Overcooked task, players cooked and delivered onion soup while seeing into another kitchen where a second player had to walk farther to do the same tasks. Participants did not know the second player was a bot asking for help. People in the Latino group gave away more onions, and the agent trained on that data replicated the tendency. In a second test, that agent also donated more money to someone in need.
The senior authors note that values should not be hard coded into AI and that demonstrations might scale if developers add more culture-specific data and fine-tune systems before deployment. Additional research is needed to test more cultural groups, competing values and real-world problems. Additional coauthors are from UW and San Diego State University; the source is the University of Washington.
Difficult words
- inverse reinforcement learning — a method that learns goals by watching actions
- altruism — willingness to help others without personal gain
- infer — to decide a fact from observed evidence or actionsinfers
- replicate — to copy or reproduce something previously observedreplicated
- demonstration — examples that show how people behavedemonstrations
- deployment — the act of making a system available for use
- cultural — relating to the customs and values of a group
Tip: hover, focus or tap highlighted words in the article to see quick definitions while you read or listen.
Discussion questions
- Do you think AI should learn cultural values by observing people rather than having values hard coded? Why or why not?
- What are possible risks and benefits of fine-tuning AI with culture-specific data before deployment?
- How could researchers test whether an AI has learned different cultural values accurately and fairly?
Related articles
After-work invitations can help some employees but harm others
New research shows after-work invitations often make socially confident employees feel connected, while shy workers can feel pressure and anxiety. Authors advise people to know their limits and for coworkers to think before inviting.
Egyptian university and pharma join to create Africa’s first biotechnology academy
The American University in Cairo and Minapharm have formed a partnership to set up what the university calls the first African academy for biotechnology. The initiative starts early this year to strengthen education, research and industry links.
Why Many Moroccan Students Are Studying in China
Many Moroccan students choose to study in China because economic and social pressures at home make staying difficult. High youth unemployment, low graduate pay, social media and agencies drive the trend, but visas and scams create risks.