UW Researchers Map Cultural Values into AI Learning Systems

Nigini Oliveira, Jasmine Li, Koosha Khalvati, Rodolfo Cortes Barragan, Katharina Reinecke, Andrew N. Meltzoff, and Rajesh P. N. Rao are developing culturally-attuned artificial intelligence (AI) systems capable of learning human cultural values. Based at the Paul G. Allen School of Computer Science and Engineering and the Institute for Learning & Brain Sciences at the University of Washington, their work proposes utilizing inverse reinforcement learning (IRL) to enable AI agents to acquire cultural norms through observation and interaction. This approach addresses the challenge of constructing a universal moral code for AI by allowing systems to adapt to specific cultural contexts, mirroring how a child learns values within a given culture. Culturally-Attuned AI and Value Learning Researchers propose a new approach to AI development: rather than creating a universal moral code, AI agents should learn values by being “embedded” in a specific human culture. This acknowledges that values, norms, and behaviors vary across cultures, necessitating AI systems capable of adapting to local tendencies.

The team views culture as a dynamic cycle of shared values, requiring AI to handle such fluidity and recognize that behaviors acceptable in one culture may not be in another.

This research focuses on teaching AI “altruistic behavior”—acting to benefit others, even at personal cost—as a culturally-relevant value.

The team posits that AI can learn this by observing human interactions, recognizing that altruism is a complex behavior shaped by cultural norms. They aim to move away from “one-size-fits-all” AI systems built on broad internet data, toward AI agents that can understand and adapt to specific cultural contexts. The methodology employs inverse reinforcement learning (IRL) to allow AI agents to implicitly acquire culturally-relevant values from human behavior. Through an online game, the researchers tested if AI could learn variations in altruistic behavior between different cultural groups, and generalize these learned values to new situations. This proof-of-concept demonstration suggests AI can be “raised” within a culture to learn typical behaviors and values through observation. The Challenge of Universal AI Values A significant challenge in AI development lies in equipping agents with culturally relevant values and behaviors. Current AI systems, often built on large-language models, tend towards a “one-size-fits-all” approach, failing to account for the diverse values across human cultures. Researchers propose that, instead of programming a universal moral code, AI should learn values by being embedded within a specific human culture, adapting to its norms and behaviors—much like a child learns from its upbringing.

This research focuses on enabling AI to learn altruism – benefitting others at a potential cost to oneself – as a key cultural value. The authors suggest that defining a universal code of values is difficult, therefore AI agents should implicitly learn them through observation. By “raising” an AI within a particular culture, it can adapt to the values and behavioral tendencies of that group, acknowledging that norms differ significantly across cultures and necessitating flexible AI systems. The proposed method utilizes inverse reinforcement learning (IRL) to allow AI agents to learn culturally specific reward functions. This approach enables an AI to acquire altruistic characteristics reflective of a particular human cultural group’s average behavior and generalize those learnings to new situations requiring altruistic judgment. This work provides a proof-of-concept for AI agents learning behaviors and values directly from observing human interaction, moving beyond generalized AI models. The end AI innocence: Genie is out of the bottle. Gill AI Learning from Human Cultural Context A key challenge in AI development is imbuing agents with culturally relevant values and behaviors. Researchers argue against a “one-size-fits-all” approach, suggesting AI should learn these values through immersion in the human culture it operates within. This recognizes the dynamic nature of culture – the shared values, norms, and behaviors of a group – and the need for AI to adapt to varying cultural expectations rather than relying on a pre-defined, universal code. The research proposes using inverse reinforcement learning (IRL) as a method for AI to implicitly learn cultural values. Specifically, the study focuses on altruistic behavior – actions benefiting others at a personal cost – as a culturally significant value. By observing human behavior, an AI agent can learn the reward functions that govern actions, ultimately reflecting the altruistic characteristics typical of a specific cultural group. This approach aims to create AI capable of handling cultural dynamics, acknowledging that behaviors considered acceptable in one culture may differ in another. The researchers highlight the importance of AI adapting to the values and tendencies of the culture in which it is deployed, recognizing that even humans struggle to fully understand and appreciate different cultural norms, hinting at the potential for AI to surpass human flexibility in this area. Man is to computer programmer as woman is to homemaker? Defining Culture for AI Applications A key challenge in AI development is equipping agents with values and behaviors reflective of human cultures. Current AI systems, often based on large-language models, risk creating a “one-size-fits-all” approach rather than acknowledging diverse cultural norms. Researchers propose that AI should learn cultural values through immersion, similar to how a child acquires them, allowing for adaptation to specific cultural contexts and recognizing that behaviors acceptable in one culture may not be in another. The research focuses on enabling AI to learn “altruistic behavior”—actions benefitting others at a personal cost—as a culturally-relevant value. The authors posit that AI agents should be able to implicitly learn these values by being “embedded” in a human culture, rather than relying on a pre-programmed universal code. This approach acknowledges the dynamic nature of culture—a continuous cycle of adopting and shaping values—and necessitates AI capable of adapting to local behavioral tendencies. To test this, researchers utilized inverse reinforcement learning (IRL) to allow AI agents to learn reward functions reflecting altruistic tendencies observed in human subjects from different cultural groups. This experimental paradigm involved an online game requiring real-time decisions, with the goal of demonstrating that AI can acquire and generalize culturally-typical behaviors and values through observation of human interactions. The study received funding from the Templeton World Charity Foundation, NSF, and other organizations. Implicit Learning of Cultural Values Researchers propose a new approach to AI development: instead of programming a universal moral code, AI agents should learn cultural values through observation, mirroring how a child learns within a specific culture. This recognizes that values, norms, and behaviors differ across groups, necessitating AI systems capable of adapting to the specific culture in which they operate. The study acknowledges the dynamic nature of culture, a continuous cycle of shaping and adopting values, and the challenges even humans face in understanding differing cultural norms.

This research focuses on “altruistic behavior” – actions benefiting others at a personal cost – as a key cultural value for AI to learn.

The team proposes using inverse reinforcement learning (IRL) as a method for AI agents to acquire these values implicitly from humans. IRL allows the AI to learn reward functions—governing its actions—by observing variations in human behavior. This allows AI to move beyond a “one-size-fits-all” approach, common in current large-language model AI systems. The study tested this approach using an experimental paradigm where AI agents learned from human subjects from two cultural groups in an online game.

Results demonstrated the AI could acquire altruistic characteristics reflective of the average behavior of each group, and generalize to new scenarios. This proof-of-concept suggests AI agents can be endowed with the ability to learn culturally-typical behaviors and values directly from observing human interaction. Altruism as a Key Cultural Value Researchers propose a shift in AI development, moving away from universal moral codes towards culturally-attuned systems. The core idea is that AI should learn values by being “embedded” in a human culture, much like a child. This approach acknowledges that values, norms, and behaviors differ across social groups and that a “one-size-fits-all” AI is insufficient. The work highlights the need for AI capable of adapting to the specific values of the culture in which it operates, recognizing inherent challenges in cross-cultural understanding.

This research specifically investigates how AI can learn “altruistic behavior”—actions benefitting others at a potential cost to oneself. Recognizing altruism as a culturally-influenced value, the team proposes using inverse reinforcement learning (IRL). IRL allows AI agents to learn reward functions – essentially, what constitutes ‘good’ behavior – by observing human actions. This learning process aims to equip AI with culturally-typical behaviors and values through direct observation of human interactions, rather than pre-programmed rules. The study focuses on enabling AI to learn from variations in altruistic behavior between cultural groups within an online game. By using IRL, the AI agents can acquire reward functions reflective of the average behavior within each group, and then generalize these learned values to new scenarios requiring altruistic judgment. This demonstrates a proof-of-concept for AI agents to implicitly acquire culturally-relevant values directly from observing human behavior, acknowledging the dynamic nature of culture itself.

Inverse Reinforcement Learning (IRL) for AI Researchers are exploring how to create culturally-attuned AI, recognizing that a universal moral code is challenging due to differing human values. The core idea is that AI should learn values by being “embedded” in a culture, much like a child. This contrasts with current AI systems based on large language models, which can result in a “one-size-fits-all” approach. The work proposes that AI agents should implicitly learn cultural values and behaviors from observation, adapting to the specific group in which they operate.

Inverse Reinforcement Learning (IRL) is proposed as a method for AI to acquire these culturally relevant values. This approach allows agents to learn reward functions – governing their actions – by observing variations in human behavior. Specifically, the research focuses on altruism – benefiting others at a potential cost to oneself – as a key cultural value for AI to learn. The goal is to enable AI to adapt to the values and behavioral tendencies of a given culture. The study tested this approach using an experimental paradigm, where AI agents used IRL to learn from the altruistic behaviors of human subjects from two cultural groups in an online game.

Results demonstrated that an AI agent could acquire characteristics reflective of a specific human cultural group and generalize these learnings to new scenarios requiring altruistic judgment. This provides proof-of-concept for AI learning culturally-typical behaviors directly from observing humans. IRL and Learning Reward Functions Researchers propose a new approach to AI development: instead of programming a universal moral code, AI agents should learn cultural values through immersion, much like a child. This recognizes that values, norms, and behaviors differ across cultures, and AI needs to adapt accordingly. The study focuses on “culture” as a dynamic cycle of shared values, acknowledging its ever-changing nature and the need for AI to respond to these shifts, rather than impose a rigid, pre-defined framework.

This research explores using inverse reinforcement learning (IRL) to enable AI to implicitly learn cultural values from human behavior. Specifically, the study investigates altruism – benefiting others at a potential cost to oneself – as a key cultural value. By observing human actions, AI agents can learn reward functions that reflect these altruistic tendencies, allowing them to make culturally-relevant decisions and adapt to the specific norms of the group they are interacting with. The researchers tested this approach using an experimental paradigm where AI agents learned from variations in altruistic behavior from two cultural groups within an online game. Results indicate that an AI agent can acquire the altruistic characteristics of a particular cultural group and generalize those learnings to new situations. This demonstrates a proof-of-concept for AI agents learning and embodying culturally-typical behaviors directly from observing human interactions. AI Learning from Altruistic Behavior Researchers are exploring how AI can learn culturally-specific values, moving beyond “one-size-fits-all” systems. The core idea is that AI should adapt to the values of the culture it’s operating within, much like a child learns from its upbringing. This approach acknowledges the dynamic nature of culture—the shared values, norms, and behaviors of a social group—and the need for AI to recognize that acceptable behaviors vary across cultures. The study proposes using inverse reinforcement learning (IRL) to allow AI agents to implicitly learn these cultural values by observing human behavior. IRL enables the AI to infer the reward functions—the underlying motivations—driving human actions. This was tested by examining altruistic behavior—actions benefiting others at a cost—in two cultural groups through an online game requiring real-time decisions, demonstrating that AI can learn culturally-typical behaviors. Ultimately, the research provides a proof-of-concept that AI can be “raised” within a culture and acquire its values directly from observing human interactions. The goal is to create AI agents capable of adapting to the values and behavioral tendencies of a specific culture, recognizing that what is considered normative varies significantly across the globe, and potentially surpassing human flexibility in this area. Testing AI with Cultural Variations Researchers are exploring how to build AI that adapts to different cultures, recognizing that a “one-size-fits-all” approach doesn’t work for values and behaviors. Instead of programming a universal moral code, they propose AI should learn cultural norms through observation, much like a child is raised within a specific culture. This approach acknowledges the dynamic nature of culture – the continuous cycle of shaping and adopting values – and the need for AI to adjust to local behavioral tendencies. The study focuses on teaching AI about altruism – acting to benefit others, even at personal cost – as a culturally-influenced behavior. The researchers hypothesize that AI can implicitly learn these values by being “embedded” in a human culture, observing and adapting to its norms. This is a departure from current AI systems, often based on large language models trained on broad internet data, which may not reflect the nuances of specific cultural groups. Using inverse reinforcement learning (IRL), the researchers demonstrated that AI agents could learn altruistic characteristics reflective of specific human cultural groups. The AI learned by observing human behavior in an online game, and could then generalize those learned values to new situations requiring altruistic judgments. This suggests a proof-of-concept for AI that can adapt its behavior based on the culture in which it operates. Large AI models are cultural and social technologies.

Generalizing Altruistic Judgments in AI Researchers are exploring how AI can adapt to differing cultural values, moving away from “one-size-fits-all” AI systems. The core idea is that AI agents should learn values by observing human behavior within a specific culture, similar to how a child acquires cultural norms. This approach acknowledges that values are dynamic and vary between social groups, necessitating AI capable of adapting to local behavioral tendencies, rather than relying on a pre-programmed universal code of ethics. The study proposes using inverse reinforcement learning (IRL) as a method for AI to implicitly learn cultural values. Specifically, researchers tested this with altruistic behavior, observing how AI agents could learn differing reward functions based on variations in human behavior from two cultural groups playing an online game. The goal was to demonstrate that an AI could acquire and generalize altruistic characteristics reflective of the group it learned from, adapting to culturally-specific decision-making. The research provides a proof-of-concept showing AI agents can learn culturally-typical behaviors through observation. Funding for this work came from multiple sources, including the Templeton World Charity Foundation, the National Science Foundation, and the Bezos Family Foundation, with no influence on the study’s design or analysis. Data and code used in the study are publicly available, enabling replication and further investigation into culturally-attuned AI development. Proof-of-Concept for Cultural AI Researchers demonstrated a proof-of-concept for “culturally-attuned AI” by leveraging inverse reinforcement learning (IRL). The goal was to enable AI agents to learn culturally specific values, rather than relying on a universal moral code. This approach proposes that AI should implicitly learn values by being “embedded” in a human culture, mirroring how a child learns from its upbringing. The study focused specifically on learning altruistic behaviors, recognizing this as a key cultural value that varies between groups. The research team tested their method using an experimental paradigm where AI agents learned different reward functions through IRL. These functions governed the agents’ actions based on observing variations in altruistic behavior from human subjects belonging to two distinct cultural groups within an online game. Results showed the AI could acquire characteristics reflective of the average behavior of each group, and generalize to new scenarios requiring altruistic judgments—demonstrating a capacity to learn culturally-typical behaviors. This work addresses the challenge of creating AI that understands and adapts to diverse cultural norms. Current AI systems often rely on large-language models offering a “one-size-fits-all” approach, but this research proposes a more nuanced method. By learning from human behavior within a specific culture, AI agents can move beyond a universal standard and potentially deal with the dynamic values and behaviors of different social groups—acknowledging that norms vary across cultures. Funding Sources for the Research Funding for this research was provided by multiple sources, demonstrating a collaborative approach to investigating culturally-attuned AI.

The Templeton World Charity Foundation supported the work of Rajesh P.N. Rao and Andrew N. Meltzoff. Additionally, the National Science Foundation provided grants—number 2223495 for RPNR and number 2230466 for Katharina Reinecke and Andrew N. Meltzoff—furthering the project’s scope. Further financial support came from the University of Washington + Amazon Science Hub, and a grant from the Bezos Family Foundation, both supporting RPNR’s work.

The Cherng Jia & Elizabeth Yun Hwang Professorship also contributed to the funding of Rajesh P.N. Rao’s research. This multi-faceted funding strategy allowed for a comprehensive exploration of implicit learning of altruistic cultural values. Importantly, the source explicitly states that the funders played no role in the study design, data collection and analysis, the decision to publish, or manuscript preparation. This highlights the research team’s independence in pursuing their investigation into how AI agents can learn culturally-typical behaviors and values from observing human interactions. Data and Code Availability The researchers developed a method for AI to learn cultural values implicitly through inverse reinforcement learning (IRL). This approach allows AI agents to acquire values by observing human behavior within a specific culture, rather than relying on a pre-programmed, universal code. Their work addresses the limitations of current AI systems built on large language models, which often lack cultural sensitivity and offer a “one-size-fits-all” solution, potentially overlooking the diverse values across different human groups. This study specifically focused on altruism—benefiting others at a potential cost to oneself—as a key cultural value for AI to learn.

The team tested their approach using an experimental paradigm where AI agents learned from variations in altruistic behavior demonstrated by human subjects from two different cultural groups within an online game. This allowed them to observe how AI agents could adapt their actions based on observed cultural tendencies, demonstrating a proof-of-concept for culturally-attuned AI. Code and data used to replicate the findings of this research are publicly available at https://github.com/JasmineLi-805/moral-ai-irl. Funding for the project was provided by the Templeton World Charity Foundation, the National Science Foundation, a UW + Amazon Science Hub grant, and the Bezos Family Foundation, with no influence on the study’s design, data analysis, or publication. The researchers emphasize the importance of AI adapting to local values rather than imposing universal standards.

Competing Interests Disclosure The authors address a key challenge in AI: imbuing agents with culturally appropriate values and behaviors. Current AI systems, often built on large-language models, tend toward a “one-size-fits-all” approach, lacking the nuance of human cultural diversity.

This research proposes that AI should learn cultural values through immersion, much like a child, adapting to the norms of the community in which it operates. This approach acknowledges that acceptable behavior varies across cultures, requiring flexible AI systems. This study focuses on altruism—benefiting others at a personal cost—as a core cultural value for AI to acquire. Recognizing that altruistic tendencies differ between cultures, the researchers utilized inverse reinforcement learning (IRL) to enable AI agents to learn from variations in human behavior. Specifically, they examined how AI could learn differing altruistic characteristics from two cultural groups through an online game involving real-time decision-making. The authors explicitly state that “the authors have declared that no competing interests exist.” Funding for this research was provided by the Templeton World Charity Foundation, the National Science Foundation, a UW + Amazon Science Hub grant, and the Bezos Family Foundation, however, these funders played no role in the study’s design, data analysis, or publication decisions. Data and code used in the study are publicly available on GitHub. Source: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0337914 Tags: