General Social Agents

Key Takeaways

• AI agents use theory‑grounded language instructions.
• Trained on small seed games, predict across 883k games.
• Outperform cognitive hierarchy and equilibrium models in tests.
• Reduce need for extensive setting‑specific data.
• Enable faster policy and market simulations.

Pulse Analysis

Applying established social‑science theories to new contexts has long required ad‑hoc tweaks and fresh data, limiting their predictive power. Recent advances in large‑language models now allow researchers to embed theoretical constructs directly into AI agents via natural‑language prompts. This approach sidesteps the traditional bottleneck of gathering setting‑specific datasets, offering a more universal toolkit for behavioral prediction across diverse environments.

Manning and Horton’s experiment leverages a modest library of "seed" games—structurally distinct but conceptually related—to train agents that can extrapolate to an expansive set of 883,320 novel games. By combining theory‑grounded instructions with empirical observations from the seed games, the agents generate simulations that mirror initial human choices. In a preregistered test covering 1,500 randomly selected games, these agents consistently outperformed classic cognitive‑hierarchy frameworks, equilibrium calculations, and generic AI baselines, demonstrating the practical strength of theory‑infused simulation.

For businesses and policymakers, this breakthrough translates into faster, cheaper scenario analysis. Companies can model consumer responses to product variations or market shocks without launching costly pilots, while regulators can anticipate behavioral outcomes of new policies in real time. As AI agents become more adept at internalizing social‑science theory, the frontier of predictive analytics expands, promising richer insights for strategic decision‑making across sectors.