Robot hand offering flowers to human hand symbolizing AI and Altruism

AI vs. Altruism: Can Language Models Predict Our Generosity?

Theo Raines monogram
Theo Raines in Tech & Innovation October 2026 • 4 min read.

"New research reveals how GPT-4 struggles to understand the balance between self-interest and the desire to help others, challenging its role in social decision-making."


Artificial intelligence is rapidly transforming many areas of our lives, from everyday tasks to high-stakes decisions. Generative AI, in particular, offers exciting possibilities for analyzing data and making informed choices. But can AI truly understand the complexities of human nature, especially when it comes to balancing our own needs with the interests of others?

A new study by Valerio Capraro, Roberto Di Paolo, and Veronica Pizziol, has investigated how well advanced AI models can predict human behavior in scenarios that require navigating self-interest and altruism. The researchers focused on a classic economics experiment called the 'dictator game,' where participants decide how to split a sum of money with a stranger. The results reveal surprising limitations in AI's ability to accurately predict our generous impulses.

This research has significant implications for how we use AI in social contexts. If AI models overestimate or underestimate human altruism, it could lead to flawed decisions in areas like public policy, business strategy, and even our expectations of each other. Understanding these biases is crucial for building AI systems that are both intelligent and socially responsible.

AI Search Multiple angles on this topic

Measuring Generosity

The impact of language models on generosity remains an open research question. Available evidence may indicate patterns in how models respond to charitable or prosocial scenarios, but such patterns should not automatically be treated as predictions of real human behavior. Any assessment should distinguish between model-generated responses and generosity expressed in real-world settings.

Current AI Approaches

Current AI systems commonly use generative assistants to answer questions, support writing, plan activities, brainstorm ideas, create images, complete work, and write code. IBM describes artificial intelligence as technology that enables machines to simulate learning, comprehension, problem-solving, decision-making, creativity, and autonomy. OpenAI frames its research as potentially leading to artificial general intelligence capable of solving human-level problems, while Google presents Gemini as an AI assistant for everyday tasks. These capabilities can support experiments about generosity, but they do not by themselves establish that a model understands or can reliably predict human altruism.

From Simulation to Exploration

Google AI Studio illustrates a contemporary stage in AI development in which users can generate photorealistic window views based on live weather and specific locations. The same platform describes managing a virtual metropolis and fulfilling tasks provided by Gemini. These examples show a shift toward interactive systems that simulate environments and carry out directed activities, providing a technological context for studying how AI might model social decisions. They do not, on their own, establish a historical milestone in predicting human generosity.

The Dictator Game: A Window into Human Altruism

Robot hand offering flowers to human hand symbolizing AI and Altruism

The dictator game is deceptively simple. One person, the 'dictator,' receives a sum of money and decides how much to share with another participant. The receiver has no say in the matter, making it a direct measure of the dictator's willingness to prioritize another person's well-being over their own.

Researchers have used the dictator game for decades to study factors influencing altruism, such as fairness, empathy, and social norms. The study leveraged data from 108 dictator game experiments conducted across 12 countries, providing a rich dataset of human behavior for AI models to analyze. This extensive dataset provided a rigorous test for evaluating the AI's predictive capabilities.

  • Standard Dictator Game: Participants can give any amount of their money to the other person.
  • Dictator Game with a 'Take' Option: Dictators start with more money and can take from the receiver.
  • Extreme Dictator Game: Participants can only choose to give nothing or everything.
AI Search Multiple angles on this topic

An Emerging Research Area

Research on whether language models can predict generosity appears to remain an emerging area rather than a settled field. Results may depend on how generosity is defined, how scenarios are presented, and whether model outputs are compared with actual human behavior. Reviews should therefore treat apparent predictive success cautiously and separate language performance from demonstrated social understanding.

Limits of Prediction

A language model can produce a plausible answer about altruism without possessing personal motives or experiencing compassion. Its response may reflect wording, training patterns, or the structure of a prompt rather than a dependable prediction of what a person will do. Consequently, confident-sounding outputs should not be treated as conclusive evidence about generosity.

Model Versus Person

Comparing a language model with human participants requires care because the two produce responses under different conditions. Models generate text from prompts, whereas people make decisions within personal, social, and material circumstances. A useful comparison should therefore examine both response quality and correspondence with real-world generous behavior, rather than treating fluent language as equivalent to human altruism.

The research team fed the experimental instructions from these games into three advanced AI chatbots: GPT-4, Bard, and Bing Chat. They then asked the chatbots to predict how people would behave, estimating the distribution of choices in each scenario. By comparing the AI's predictions with the actual human behavior, the researchers could assess how well the models understood the delicate balance between self-interest and the interest of others.

What This Means for the Future of AI

The study underscores the need for careful consideration of AI's limitations, especially when it comes to understanding complex human motivations like altruism. As AI takes on a greater role in shaping our social policies and decisions, it's crucial to be aware of its potential biases and to develop methods for mitigating these issues. While generative AI holds immense potential, its effectiveness hinges on its ability to accurately reflect the nuances of human behavior, ensuring that technology serves to enhance, rather than distort, our understanding of one another.

AI Search Multiple angles on this topic

Interpreting AI Altruism

Language models may offer useful tools for exploring how generosity is described, reasoned about, or represented in text. That usefulness does not prove that they possess altruistic motives or can accurately forecast individual decisions. The strongest conclusions should remain proportionate to the evidence and should distinguish simulated reasoning from human moral action.

Testing Social Prediction

Future work could examine whether language models' predictions remain reliable across different generosity scenarios and populations. It could also compare model outputs with observed decisions instead of relying only on hypothetical answers. Progress will depend on clearer definitions, transparent evaluation, and careful attention to the gap between prediction and intervention.

AI for Collective Benefit

Google describes its AI work as aiming to enrich knowledge, solve complex challenges, and help people grow through useful AI tools and technologies. That broad goal places generosity research within a larger question about how AI systems should serve people and communities. Even when tools are designed to be helpful, their social value depends on how they are developed, evaluated, and used.

Human Choices Matter

Generosity is expressed through human choices, relationships, circumstances, and consequences that a text-based system may only partially represent. A model can help people discuss or organize prosocial actions, but the real-world impact ultimately depends on human judgment and participation. Claims about AI predicting altruism should therefore remain attentive to the people affected by those predictions.

About this Article -

Written with AI assistance from published research, and reviewed by the Mystum team. See our About page for more information.

Everything You Need To Know

1

What is the 'dictator game' and how does it help researchers study altruism?

The 'dictator game' is a simple economics experiment where one person, the 'dictator,' decides how to split a sum of money with another participant. The receiver has no say, making it a direct measure of the dictator's willingness to prioritize another person's well-being over their own. Researchers use this game to study altruism because it isolates the decision-maker's inherent generosity. It allows them to observe how individuals balance their self-interest against the needs of others, providing insights into factors like fairness, empathy, and social norms that influence altruistic behavior. The study mentioned in the context used the 'dictator game' to test how well AI models such as GPT-4, Bard, and Bing Chat, could predict human behavior in scenarios involving altruism.

2

How did researchers evaluate AI's understanding of human altruism in the study?

The researchers evaluated AI's understanding of human altruism by feeding the experimental instructions from different versions of the 'dictator game' into advanced AI chatbots: GPT-4, Bard, and Bing Chat. They then asked the chatbots to predict how people would behave in these games, specifically estimating the distribution of choices in each scenario. The team used data from 108 'dictator game' experiments across 12 countries to give the AI models a broad data set. By comparing the AI's predictions with the actual human behavior observed in those experiments, the researchers could assess how accurately the AI models understood the balance between self-interest and the interest of others. This comparison revealed the limitations of the AI models in predicting human generosity.

3

What are the different variations of the 'dictator game' used in the study, and how do they differ?

The study utilized three variations of the 'dictator game' to provide a comprehensive test of AI's understanding of human altruism. The first, the 'Standard Dictator Game,' allowed participants to give any amount of their money to another person. The second, the 'Dictator Game with a 'Take' Option,' gave dictators more money initially and the ability to take money from the receiver. This introduced a layer of potential self-interest that the AI needed to navigate. Finally, the 'Extreme Dictator Game' offered a stark choice: participants could only give either everything or nothing. Each variation presented different decision-making scenarios, allowing researchers to evaluate how AI models like GPT-4, Bard, and Bing Chat, responded to varying degrees of complexity in the context of altruistic choices.

4

What implications does this research have for the future use of AI in social contexts?

This research has significant implications for the future use of AI in social contexts. The study's findings that AI models like GPT-4, Bard, and Bing Chat, struggle to accurately predict altruistic behavior raise concerns about using AI in areas like public policy, business strategy, and social decision-making. If AI systems overestimate or underestimate human generosity, it could lead to flawed decisions. To mitigate these issues, it's crucial to be aware of AI's potential biases and to develop methods for improving AI's understanding of complex human motivations. This involves ensuring that AI systems are trained on diverse datasets and are constantly evaluated and refined to accurately reflect the nuances of human behavior, so that technology enhances, rather than distorts, our understanding of one another.

5

How can the limitations of AI, as demonstrated in this study, be addressed to improve its ability to understand human behavior?

Addressing the limitations of AI, as demonstrated in its struggles to understand human altruism, requires a multifaceted approach. One key area is improving the data used to train these AI models. Training AI models like GPT-4, Bard, and Bing Chat, on more extensive and diverse datasets of human behavior, particularly those reflecting real-world interactions in the 'dictator game,' is essential. Another vital aspect is refining the AI algorithms themselves, to better capture the complexities of human motivations, including empathy, fairness, and social norms. Continuous evaluation and validation of AI models against real-world human behavior are also necessary. Furthermore, integrating ethical considerations into the development and deployment of AI, ensuring that these systems are designed to be socially responsible, will be critical for building AI systems that can accurately predict human behavior and contribute positively to society.

Newsletter Subscribe

Subscribe to get the latest articles and insights directly in your inbox.