Artificially intelligent agents in the social and behavioral sciences: A history and outlook

TL;DR

The paper traces a 75-year history of agentic AI in the social and behavioral sciences, from early computer simulations to modern generative AI, emphasizing bidirectional influence between AI innovations and scientific practice. It argues that AI serves as both a tool and a subject of study, shaping methods, theories, and ethical considerations, while social sciences, in turn, guide AI development. The work maps the progression through simulations, cybernetics, complexity science, big data, and GenAI, highlighting methodological shifts toward prediction, causal inference, and human-AI interactions. It concludes with a forward-looking outlook on how human-AI co-evolution will redefine knowledge production, with attention to replication, governance, and epistemic responsibilities.

Abstract

We review the historical development and current trends of artificially intelligent agents (agentic AI) in the social and behavioral sciences: from the first programmable computers, and social simulations soon thereafter, to today's experiments with large language models. This overview emphasizes the role of AI in the scientific process and the changes brought about, both through technological advancements and the broader evolution of science from around 1950 to the present. Some of the specific points we cover include: the challenges of presenting the first social simulation studies to a world unaware of computers, the rise of social systems science, intelligent game theoretic agents, the age of big data and the epistemic upheaval in its wake, and the current enthusiasm around applications of generative AI, and many other topics. A pervasive theme is how deeply entwined we are with the technologies we use to understand ourselves.

Figures (5)

• Figure 1: An illustration of the Sugarscape model by Epstein and Axtell sugarscape---an archetypal agent-based model. The simulation is based on agents spread out over a grid and an unevenly distributed natural resource---the "sugar." A darker shade of the background color indicates a higher sugar concentration. In principle, one could use any grid geometry and sugar distribution, but Sugarscape has become synonymous with the shown bimodal distribution on a 50 $\times$ 50 grid. Agents move to optimize a utility function within their "field of vision"---the gray surroundings of the white dots in panel A. If direct sugar consumption is all the agents care about, their population distribution will come to resemble the sugar distribution (panel A). By adding abilities to the agents and terms to their utility functions, the Sugarscape simulation can recreate many spatial features and functions of a society and economy, including the emergence of ethnicities, borders, border trade, a segregation of people by their aversion toward pollution, etc. (panel B). It is also possible to extend the simulation with non-local relationships between agents, to make Sugarscape the basis of a social network simulation.
• Figure 2: Example output from network models of agents trying to optimize their network positions, adapted from Refs. bala_goyal and diplo. Panel A illustrates an economics-style model where agents optimize a utility function that represents a trade-off between the number of agents whose information can reach them (directly or indirectly through paths of directed links) and the in-degree (i.e., the number of direct-link neighbors) bala_goyal. The figure shows agents updating their links until a Nash equilibrium (which is an absorbing state) is reached (at time step 11). The updating proceeds by agents either configuring their network to what would have been optimal in the previous time step (green arrows) or, with some probability, doing nothing (blue arrows). Panel B presents snapshots of a model with agents again attempting to optimize conflicting objectives: namely, simultaneously maximizing their closeness centrality (a proxy for power) and minimizing their degree (the cost of communication) ghoshal_holme. The agents can reconfigure their position locally (within their second neighborhoods) and use reinforcement learning based on their own and their neighbors' performance to decide their updating strategy. The updating strategies are codified as a priority list of actions to take (e.g., deleting the link to the lowest-degree neighbor). The symbols represent actions of highest priority for the respective agent.
• Figure 3: The DARPA Red Balloon challenge was a contest in social mobilization for geo-location to commemorate the Internet's 40th anniversary. On December 5, 2009, ten red balloons were positioned at undisclosed locations across the continental USA (panel A). The first participant to correctly report the location of all was awarded $40,000. The problem was designed to exploit (then) new technologies like social media and crowdsourcing to scale up the recruitment and organization of human work. Panel B illustrates the winning team's strategy for incentivizing team-member recruitment inspired by Ref. dodds2003experimental. Discovering a balloon earned a $2000 share of the total prize sum, while recruiting someone earned half of that person's award tang2011reflecting.
• Figure 4: An illustration of the LLM-chatbot-driven simulations of Ref. bernstein where 25 agents simulated the inhabitants of a village. Each agent was seeded by a description of their background and relation to other agents. The authors observed several aspects of emergent sociality, such as the coordinated organization of a Valentine's Day party. (Reprinted with permission from the authors.)
• Figure 5: A timeline of some key moments covered in this paper.