A study by the University of Washington in the United States has shown for the first time that artificial intelligence (AI) systems can learn and internalize corresponding cultural values by observing human behavior in specific cultures. This provides a new approach to solving the problem of cross-cultural adaptation in AI. The relevant paper has been published in PLOS One. At present, AI usually conducts training based on large-scale Internet data, and the values contained in these data often have cultural bias, resulting in inconsistent performance of the system in front of users with different cultural backgrounds. To this end, the research team attempted to enable AI to learn values from the behavior of specific cultural groups through "observational learning," rather than being pre implanted with a set of universal guidelines. The research team is trying to explore whether AI can naturally learn the values of its culture by observing the behavior of people around it, just like children do. The team recruited 190 adults to participate in the experiment and had them interact with AI agents separately. In the experiment, participants played a collaborative task adapted from the game "overcooked". In the game, participants can choose whether to donate their acquired resources to robot players who are clearly at a disadvantage for free, although this will affect their task scores. The results showed that a group of participants exhibited more altruistic behavior overall. AI agents use the "reverse reinforcement learning" method to infer the behavioral goals and intrinsic values of the observed group from their behavior. In subsequent testing, these agents successfully extended their learned "altruistic tendencies" to new scenarios such as donating funds: AI trained on this human group data showed higher generosity in donation tasks. This is similar to the way children learn - they are not repeatedly trained to do something, but learn social behaviors such as sharing and caring through observing their parents' interactions with others. ”The co-author of the paper, Andrew Melzoff, a psychology professor at the University of Washington, stated that "values are more 'captured' than 'taught'." The team believes that creating AI with cultural adaptability and understanding of others' perspectives is an important issue facing society today. With the increasing cultural diversity and volume of input data, such methods are expected to help develop AI systems that are more tailored to specific cultural backgrounds. However, the study is still in the conceptual validation stage, and its feasibility needs to be further verified in more cultural contexts, value conflict scenarios, and complex real-world problems in the future. (New Society)