AI Brain vs Human Altruism: A depiction of the conflict between AI's predictive capabilities and the complexities of human behavior.

Can AI Really Understand Us? New Study Questions AI's Ability to Predict Human Behavior

Beau Callahan monogram
Beau Callahan in Tech & Innovation October 2026 • 4 min read.

"Groundbreaking research reveals that even the most advanced AI models struggle to accurately predict how humans balance self-interest with the needs of others, raising concerns about AI's role in social decision-making."


Artificial intelligence is rapidly transforming how we live and work, offering powerful new tools for everything from creative writing to medical consultations. As AI systems become more sophisticated, their potential to assist in complex decision-making processes is drawing increasing attention. However, many crucial decisions involve balancing our own self-interest with the interests of others, a complex equation that requires a deep understanding of human nature.

A recent study by Valerio Capraro, Roberto Di Paolo, and Veronica Pizziol sheds light on AI's current capabilities in this critical area. Their research investigates whether today's most advanced AI models can accurately predict how humans behave in situations that require balancing self-interest and the well-being of others. The findings reveal a significant gap between AI predictions and actual human behavior, raising important questions about AI's readiness for high-stakes social decision-making.

This article delves into the details of the study, exploring its methodology, key findings, and the potential implications for AI developers, policymakers, and anyone interested in the future of AI's role in society.

AI Search Multiple angles on this topic

Current Statistics & Impact

Artificial intelligence has become a transformative force across global industries, with systems capable of performing advanced tasks that once required human cognition. The technology's reach continues to expand as organizations integrate AI for data analysis, automation, and decision-support functions. While exact adoption metrics vary, AI's impact on productivity, creativity, and problem-solving represents a significant shift in how work gets accomplished.

Standard Approach, Accepted Methods & Their Limitations

Artificial intelligence is defined as a set of technologies that empower computers to learn, reason, and perform advanced tasks such as understanding language and analyzing data, bringing meaningful change to societies. Research organizations like OpenAI pursue artificial general intelligence to solve human-level problems, while platforms like Google Gemini demonstrate AI's practical assistance in writing, planning, and brainstorming. Wikipedia frames AI within the broader context of machine intelligence, acknowledging both its potential and the constraints of current systems.

Historical Perspective, Milestones, Foundational Discoveries

The field of artificial intelligence has evolved through successive waves of innovation, from early symbolic reasoning systems to modern neural networks and deep learning architectures. Key milestones include breakthroughs in machine learning algorithms, increases in computational capacity, and the development of frameworks that enable machines to recognize patterns, understand language, and solve complex problems. These foundational advances have progressively expanded AI's capabilities across diverse application domains.

How Accurate is AI in Predicting Human Altruism?

AI Brain vs Human Altruism: A depiction of the conflict between AI's predictive capabilities and the complexities of human behavior.

The researchers focused on a classic economics experiment called the "dictator game." In this game, one person (the dictator) is given a sum of money and decides how much to share with another participant. This simple scenario provides a direct measure of how people weigh their self-interest against their concern for others.

Capraro and his team presented three leading AI chatbots – GPT-4, Google's Bard, and Microsoft's Bing Chat – with experimental instructions from 108 different dictator game experiments conducted across 12 countries. The AI models were then asked to predict how people would behave in each scenario. Here's what they discovered:

  • GPT-4 Showed Qualitative Understanding: Only GPT-4 captured the basic patterns of human behavior, identifying three main groups: those primarily motivated by self-interest, those driven by fairness and aversion to inequality, and those who behaved altruistically.
  • AI Overestimates Altruism: Despite this qualitative understanding, GPT-4 consistently overestimated how generous people would be. Its predictions showed a significantly higher level of altruism than what was observed in the actual human experiments.
  • Bias Towards Idealism: The AI model underestimated self-interested and inequity-averse behaviors, suggesting an overly optimistic view of human nature.
AI Search Multiple angles on this topic

Latest Research and Reviews

Current AI research spans a wide spectrum of approaches, including large language models, multimodal systems, and specialized architectures designed for specific domains. Academic and industry publications regularly evaluate model performance, safety considerations, and emerging capabilities, while also identifying limitations such as reasoning gaps, hallucination tendencies, and data dependency. The research landscape reflects ongoing efforts to balance capability advancement with responsible deployment frameworks.

Counter Arguments and Failures

AI systems face acknowledged limitations including occasional generation of inaccurate information, biases inherited from training data, and challenges with true contextual understanding. Deployments have encountered issues ranging from unintended outputs to failures in edge-case scenarios, prompting discussions about reliability and trustworthiness. The technology's current boundaries suggest that human oversight remains essential for critical applications.

Comparative Analysis

Different AI systems exhibit varying strengths and weaknesses depending on their architecture, training data, and intended use cases. Some excels at language generation while others perform better in coding, analysis, or creative tasks. Comparative evaluations typically examine factors such as accuracy, coherence, reasoning ability, and adherence to user intent across a range of prompts and domains, highlighting that no single system currently dominates all dimensions of performance.

These findings highlight a crucial challenge: while AI can identify broad behavioral patterns, it struggles to accurately predict the nuances of human decision-making, particularly when social factors are involved. The tendency to overestimate altruism could have significant consequences if AI is used to guide decisions in areas like public policy or business strategy.

The Future of AI and Human Understanding

The study's authors suggest that the overestimation of altruism may stem from biases in AI training data or from the "reinforcement learning with human feedback" (RLHF) stage, where AI models are trained to generate human-like language. This process might inadvertently lead AI to adopt an overly positive view of human behavior in an attempt to be "politically correct."

AI Search Multiple angles on this topic

Synthesis & Expert Commentary

Discussions about AI's role and capabilities bring together diverse perspectives from technologists, ethicists, and policymakers. The consensus emerging from cross-disciplinary dialogue is that while AI demonstrates impressive functionality across many tasks, it does not possess genuine understanding or consciousness. Expert commentary often focuses on the distinction between sophisticated pattern matching and authentic comprehension, emphasizing the technology's tool-like nature.

Future Outlook & Next Frontiers

Research directions point toward more sophisticated reasoning capabilities, improved multimodal integration, and enhanced safety mechanisms for AI systems. Investigators are exploring architectures that could support more reliable generalization, reduced hallucination rates, and better alignment with human values. The trajectory suggests continued expansion of AI's functional reach while maintaining awareness of persistent technical constraints.

Broader Context & Systemic Challenges

AI's integration into societal structures raises questions about labor displacement, information quality, and the distribution of technological benefits. Systemic challenges include the need for robust governance frameworks, transparency mechanisms, and safeguards against misuse. The technology's rapid advancement outpaces many regulatory and ethical development efforts, creating alignment gaps between capability and oversight.

The Human Element & Real-World Impact

Human-AI interaction shapes how individuals and organizations incorporate these tools into daily workflows. Users report assistance with writing tasks, planning, brainstorming, and coding, while also encountering situations where AI outputs require verification or correction. The real-world impact reflects a partnership dynamic where human judgment complement machine capabilities, particularly for decisions carrying significant consequences.

How other studies frame AI predictions of human behavior

A separate benchmark likewise found a gap between describing the shape of human choices and predicting their exact distribution: GPT-4 captured three peaks qualitatively, but its numerical predictions were imprecise.[1]

Another study argues that divergence from human response patterns, including in unfamiliar scenarios, may mean apparent human-like performance sometimes reflects memorization rather than generalizable prediction.[2]

The study at a glance

Populationhuman participants from 12 countries[3]
Sample size108 experiments[3]
ComparisonGPT-4, Bard, and Bing[3]
Times cited2[3]

What these prediction results add up to

Recognizing broad kinds of behavior is not the same as estimating how common each kind is. The separate benchmark’s finding of imprecise numerical predictions illustrates that distinction.[1] [3]

The pattern across studies leaves open whether chatbot predictions reflect transferable insight into people or familiarity with examples resembling those encountered before.[2] [4]

The paper frames social decision support as requiring attention to both personal and other-regarding interests; these experiments test prediction in a specific game context, not dependable assistance across all consequential choices.[3]

What keeps these findings from being settled

  • The authors warn that optimistic predictions about generosity can distort expectations and contribute to poor policy or business choices, frustration, or conflict.[3]

What remains unresolved about prediction

  • How well would predictions represent psychological variation across populations, rather than reproduce an average that misses some groups?[5] [6]

Sources

  1. A publicly available benchmark for assessing large language models’ ability to predict how humans balance self-interest and the interest of others (pmc.ncbi.nlm.nih.gov)
  2. Take caution in using LLMs as human surrogates (pmc.ncbi.nlm.nih.gov)
  3. Assessing Large Language Models' ability to predict how humans balance self-interest and the interest of others (arXiv (Cornell University), 2023)
  4. Testing for completions that simulate altruism in early language models. (pubmed.ncbi.nlm.nih.gov)
  5. Large language models show amplified cognitive biases in moral decision-making (pmc.ncbi.nlm.nih.gov)
  6. Money, Food, and Daily Life Objects Are Similarly Shared in the Dictator Game. A Study among Poles and Tsimane’ (pmc.ncbi.nlm.nih.gov)

About this Article -

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

This article is based on research published under:

DOI-LINK: https://doi.org/10.48550/arXiv.2307.12776,

Title: Assessing Large Language Models' Ability To Predict How Humans Balance Self-Interest And The Interest Of Others

Subject: econ.gn cs.ai cs.cy cs.gt q-fin.ec

Authors: Valerio Capraro, Roberto Di Paolo, Veronica Pizziol

Published: 21-07-2023

Everything You Need To Know

1

What were the key findings of the study regarding AI's ability to predict human behavior in the dictator game?

The study, conducted by Valerio Capraro, Roberto Di Paolo, and Veronica Pizziol, used the dictator game to assess how well AI models like GPT-4, Google's Bard, and Microsoft's Bing Chat could predict human behavior. Key findings revealed that GPT-4 showed a qualitative understanding of behavior patterns, identifying groups based on self-interest, fairness, and altruism. However, the AI models, especially GPT-4, consistently overestimated altruism, and underestimated self-interested behavior. This suggests a significant gap in the AI's ability to accurately model the nuances of human decision-making when social factors are involved.

2

How did the study assess the AI models' understanding of human behavior, and what specific experiment was used?

The study evaluated the AI models by having them predict outcomes in the dictator game, a classic economics experiment. In this game, one person (the dictator) decides how to share a sum of money with another participant. The AI models, including GPT-4, Bard, and Bing Chat, were presented with instructions from 108 different dictator game experiments from 12 countries. The AI was then asked to predict how people would behave in each scenario, allowing the researchers to measure the accuracy of the AI's predictions against actual human behavior in situations that involve balancing self-interest against the well-being of others.

3

What are the potential implications of AI overestimating human altruism, as demonstrated in the research?

The overestimation of altruism by AI models, particularly GPT-4, could have significant implications. If AI systems are used to guide decisions in areas like public policy or business strategy, this bias could lead to inaccurate predictions and potentially suboptimal outcomes. For example, in resource allocation or negotiation scenarios, AI might predict more cooperative behavior than actually exists, leading to ineffective strategies. The tendency of AI to overestimate generosity highlights a critical need to improve AI's understanding of human motivations to ensure reliable decision-making support.

4

Why does AI, according to the study, struggle to accurately predict human behavior, and what factors might contribute to this?

According to the study, the challenges AI faces in predicting human behavior stem from several factors. The primary issue is the inability to accurately model the nuanced balance between self-interest and the needs of others, which is fundamental to many human decisions. The research suggests that the overestimation of altruism may result from biases in AI training data or the 'reinforcement learning with human feedback' (RLHF) stage, where AI models are trained to generate human-like language. This process might lead the AI to adopt an overly positive view of human behavior in an attempt to be 'politically correct', thus distorting the model's predictions.

5

How does the study's findings relate to the broader conversation about AI's role in society and its potential impact on decision-making processes?

The study's findings significantly influence the conversation about AI's role in society by highlighting the limitations of current AI models in understanding human behavior, especially in social contexts. As AI systems become increasingly integrated into decision-making processes, from public policy to business strategy, the ability to accurately predict human actions is crucial. The research by Capraro, Di Paolo, and Pizziol raises important questions about the reliability and ethical implications of using AI in high-stakes scenarios, particularly when decisions require an understanding of complex social dynamics. It underscores the need for further research and development to improve AI's ability to model human behavior realistically, accounting for the full spectrum of human motivations, including self-interest and altruism.

Newsletter Subscribe

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