[Paper Review] Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus?

John J. Horton, Filippas, Apostolos+1 more|arXiv (Cornell University)|Jan 18, 2023

Economic Policies and Impacts47 citations

TL;DR

The paper argues that GPT-3-like LLMs can function as implicit computational models of humans (homo silicus) and qualitatively reproduce classic behavioral economics results through simulations, enabling low-cost, scalable pilot studies before real-world experiments.

ABSTRACT

We argue that newly-developed large language models (LLMs), because of how they are trained and designed, are implicit computational models of humans -- a Homo silicus. LLMs can be used like economists use Homo economicus: they can be given endowments, information, preferences, and so on, and then their behavior can be explored in scenarios via simulation. Experiments using this approach, derived from Charness and Rabin (2002), Kahneman et al. (1986), Samuelson and Zeckhauser (1988), Oprea (2024b), and Horton (2025), show qualitatively similar results to the original, and when they differ, it is often generative for future research. We discuss potential applications, conceptual issues, and why this approach can inform the study of humans.

Motivation & Objective

• Propose and validate using LLMs as implicit models of human economic behavior (homo silicus).
• Demonstrate that GPT-3 can reproduce qualitative findings from behavioral economics experiments (e.g., dictator games, price gouging, status quo bias).
• Show how endowing LLMs with beliefs, preferences, and frames affects their choices in economic scenarios.
• Highlight the potential of in silico piloting to explore parameter spaces cheaply before real-world testing.

Proposed method

• Endow GPT-3 models with different social preferences (equity, efficiency, self-interest) and observe dictator-game choices across scenarios.
• Manipulate framing and political-like beliefs to study effects on price-gouging judgments in a Kahneman-style scenario.
• Replicate status quo bias by varying how funding allocations are framed as status quo versus neutral.
• Run a hiring scenario to examine wage and experience trade-offs under a minimum wage, analyzing observed choices and outcomes.
• Compare outputs across GPT-3 variants (text-davinci-003, text-ada-001, text-babbage-001, text-currie-001) and temperatures to assess robustness.
• Discuss the interpretation of observations given the inferential nature of LLM responses and potential performativity concerns.

Figure 1 : Charness and Rabin ( 2002 ) Simple Tests choices by model type and endowed “personality”

Experimental results

Research questions

• RQ1 Can GPT-3 style LLMs reproduce qualitative patterns from classic behavioral economics experiments?
• RQ2 How do endowments of preferences, frames, and political-like views in LLMs affect their economic choices?
• RQ3 What is the value of LLM-based simulations for piloting social science experiments in terms of cost and speed?
• RQ4 To what extent do model capabilities limit or enable behavior akin to homo economicus or humane decision-making in these tasks?
• RQ5 How robust are AI-generated results across model variants and prompt formulations?

Key findings

• LLMs can mimic inequity-, efficiency-, and self-interest–driven choices in dictator-game settings, with advanced models showing sensitivity to endowed preferences.
• In price-gouging and framing tasks, AI agents’ fairness judgments vary with price level and political framing, resembling human patterns and exhibiting framing effects in some setups.
• Status quo framing biases appear in AI responses when options are presented as the status quo, mirroring the human status quo bias.
• A minimum wage scenario induces wage increases and shifts toward more experienced hires in AI simulations, indicating labor-labor substitution effects at the simulated level.
• Capable GPT-3 models qualitatively reproduce several classic behavioral results at negligible cost, enabling rapid, large-scale exploratory studies.
• The work argues for treating LLMs as a laboratory tool for rapid, in silico experimentation to guide real-world empirical work.

Figure 2 : Kahneman et al. ( 1986 ) price gouging snow shovel question, with endowed political views