In the AI era, new research directions in behavioral science research should be explored. Photo: TUCHONG

The rapid development of artificial intelligence (AI) technologies has deeply permeated socioeconomic life and scientific research across disciplines. AI is increasingly becoming a key driver of scientific progress and industrial transformation, while also bringing significant innovations to research paradigms across a wide range of fields. Behavioral science is no exception. As an emerging technology, AI is fundamentally reshaping human decision-making environments, interaction partners, and even the mechanisms of decision-making itself, exerting far-reaching influence on every stage of the decision-making process and helping to reshape human decision-making paradigms.

More notably, as AI technologies continue to advance—particularly with the remarkable “emergent” capabilities demonstrated by large language models (LLMs) and other advanced AI systems—AI itself has begun to be viewed as an autonomous decision-making agent. This shift implies that behavioral research paradigms, theoretical frameworks, and experimental methods originally developed to study human decision-makers can now be replicated and applied to a new domain in which AI itself becomes the research subject. At the same time, researchers have found that the behavioral characteristics exhibited by rapidly evolving AI systems increasingly resemble those of humans. This observation points to a new methodological possibility: AI systems can be used as substitutes for human participants in traditional behavioral science surveys and experiments, and can therefore be utilized to generate simulated research data. A systematic examination of the new research topics and methodologies that AI brings to behavioral science is therefore necessary, both to provide researchers with practical guidance and to offer forward-looking perspectives for future inquiry.

Human attitudes and behaviors in interactions with AI

The first research domain centers on human attitudes and behaviors in interactions with AI.

Changes in human attitudes and behaviors during interactions with AI have emerged as a new area of research. Human–AI interaction can generally be categorized as either one-way interaction or two-way interaction.

In one-way interactions, AI functions as a tool used by humans and therefore occupies a passive role, with research primarily focused on human attitudes toward AI and their responses to it. In two-way interactions, however, AI can act as an autonomous agent capable of influencing human decision-making during the interaction process. The focus of research thus shifts toward whether such interactions foster human–AI collaboration and complementarity, or instead give rise to new forms of human–AI competition.

AI impact on human behavior and preferences

The second research domain emphasizes the impact of AI on human behavior and preferences.

Human behavior and preferences may also evolve through interactions with AI, with existing research suggesting that these changes can produce both beneficial and adverse effects.

On the positive side, AI can improve the accuracy of human decision-making by analyzing large volumes of data and providing decision support and optimization recommendations. In addition to improving decision quality in situations where objective standards of evaluation exist, AI can also improve individual welfare in personalized decision-making contexts by offering customized services and recommendations based on user preferences. In more complex and systemic decision-making scenarios, AI can help individuals improve both the efficiency and quality of their decisions. The development of AI technologies may also contribute to social and cultural inclusiveness by fostering greater mutual understanding among different groups.

On the negative side, several concerns have emerged. First, excessive reliance on AI-generated recommendations may lead individuals to neglect their own judgment, weakening decision-making autonomy and potentially causing certain skills to atrophy over time. Second, AI may reduce people’s emotional attentiveness to others, as well as their evaluation of others during interpersonal interactions. Third, AI systems are trained on data collected by humans, and such data inevitably contains human biases and value judgments, which may amplify algorithmic bias. Fourth, personalized recommendation systems on social media platforms may exacerbate “information cocoon” or “echo chamber” effects, contributing to greater polarization of opinion.

Decision-making characteristics of LLMs

The third research area attempts to analyze the decision-making characteristics of LLMs.

Behavioral science is no longer confined to the study of human behavior; it has increasingly begun to examine the behavioral characteristics of AI agents powered by LLMs. LLMs possess advanced natural language processing capabilities and can closely approximate human abilities in receiving and generating textual information. When viewed as decision-making agents within socioeconomic systems, LLMs raise important questions about whether their decisions differ from those of humans when they are confronted with the same text-based decision-making tasks.

At a fundamental level, LLMs discover latent patterns and regularities in massive textual datasets by learning the statistical relationships among words and texts. Through this process, they display emergent capabilities that resemble certain aspects of human reasoning and decision-making. Owing to these capabilities, research has shown that LLMs exhibit personality traits similar to humans when responding to personality questionnaires, and may even develop distinctive personalities and seemingly emotional reactions.

Behavioral scientists have begun exploring the behavior of AI agents by assigning them specific endowments, information sets, and preference structures, then observing their actions in simulated environments. Five major dimensions are commonly used to analyze the behavioral characteristics of LLMs: rationality, personality traits, theory-of-mind capabilities, behavioral preferences, and behavioral heterogeneity.

AI-based methodological innovation

The fourth research direction leads to AI-based methodological innovations in experimental research.

As growing evidence suggests that LLMs exhibit substantial similarities to humans in behavioral and decision-making characteristics, many studies have begun to explore the possibility of using LLMs as experimental subjects in place of human participants. At present, two traditional behavioral research approaches have particularly benefited from this development: survey research and micro-level experiments based on LLM-agent simulations, and macroeconomic and macrosocial experimental research based on multi-agent systems.

The basic approach to AI-assisted survey research can be summarized in two steps. First, participant profiles are constructed based on individual characteristics. Second, these identity attributes, together with survey questionnaires, are provided as inputs to LLMs. Findings from such studies point to a foundational conclusion regarding the application of LLMs in behavioral research: LLMs can serve as effective tools for simulating human beliefs. Since beliefs form the basis of human decision-making, this conclusion further suggests a novel experimental approach—conducting behavioral experiments using AI agents as substitutes for human participants.

A second methodological innovation involves multi-agent experimental research. Behavioral economics frequently examines how individuals make decisions in environments characterized by conflict or cooperation, as well as how individual behavior shapes the evolution of economic systems. Similar to traditional behavioral experiments that collect decision-making data from human subjects, multi-agent experiments treat each AI agent as an experimental participant. By clearly specifying the experimental environment, decision tasks, incentive structures, and other relevant conditions through prompt design, researchers can generate behavioral interaction and decision-making data from LLMs. This represents a new paradigm in behavioral science research—allowing AI agents to participate directly in behavioral experiments and generate experimental data.

Traditional macroeconomic theory often relies on the assumption of a representative subject, abstracting the entire economic system into a single decision-making entity. In reality, however, economic actors are heterogeneous, and their interactions give rise to complex economic systems. As a result, agent-based modeling (ABM) has gained increasing prominence as a research approach, allowing heterogeneous agents to interact under non-equilibrium conditions.

A typical ABM framework consists of three core components: agents, environments, and interaction rules. The research process generally involves three stages. First, multiple AI agents are constructed through prompt engineering and the incorporation of specialized functional modules, enabling them to play different roles within a complex economic system. Second, the economic environment is established by defining scenarios, utility functions, and interaction mechanisms among agents—often referred to as the “behavioral module.” Third, scenario simulations are conducted, and the aggregated responses of individual agents are analyzed to derive system-level economic outcomes, thereby providing insights and evidence for policy decision-making.

Yang Yang is an associate professor from the Lingnan College at Sun Yat-sen University. This article has been edited and excerpted from China Journal of Econometrics, Issue 5, 2025.