Large Language Models as Simulated Economic Agents: A New Frontier in Economics (3min read)

In the intersection of artificial intelligence and economics, a pioneering paper emerges that challenges our traditional understanding of human behavior in economic contexts. Authored by John J. Horton, the paper titled "Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus?" presents a novel approach to simulating human economic decisions using large language models (LLMs) like GPT-3. This groundbreaking work not only opens new doors in economic research but also paves the way for a deeper understanding of human nature through the lens of AI. Let's embark on a journey into the world of "homo silicus" and explore the intriguing questions and answers that this paper offers. 🚀

Homo Silicus: A New Computational Model 🧠

The paper introduces the concept of "homo silicus," a computational model of humans that can be used like economists use homo economicus. By giving LLMs endowments, information, preferences, and placing them in various scenarios, their behavior can be explored through computational simulation.

Interesting Questions and Answers 🧩

Can LLMs be Used to Simulate Human Behavior?

Answer: Yes, LLMs can respond realistically to a wide range of textual inputs, giving responses similar to what we might expect from humans. They can be used to simulate economic scenarios, providing fresh insights and qualitative results.

How Do LLMs React to Different Social Preferences?

Answer: Experiments using unilateral dictator games show that endowing the AI with various social preferences affects play. For example, instructing the AI that it cares about equity will cause it to choose equitable outcomes.

What is the Impact of Framing and Political Views?

Answer: Experiments motivated by Kahneman et al. (1986) show that large price gouging is viewed more negatively, and endowed political views matter. AI agents of the right are more sanguine about gouging generally.

Are LLMs Subject to Status Quo Bias?

Answer: Yes, GPT-3 text-davinci-003 is found to be subject to status quo bias, replicating human tendencies in decision-making scenarios.

Can LLMs Simulate Hiring Scenarios?

Answer: Yes, a hiring scenario experiment shows that the imposition of a minimum wage causes a shift in the hiring of more experienced applicants.

What is the Value of AI Experiments?

Answer: AI experiments can be used to pilot studies via simulation first, searching for novel social science insights. They offer enormous advantages in terms of speed and cost.

Are There Critiques and Limitations?

Answer: Yes, critiques include the "Garbage in, Garbage out" problem, the "performativity" problem, and concerns about what counts as an "observation." However, the paper argues that these issues can be addressed or are less significant than they might seem.

Conclusion: A New Horizon in Economics 🌅

The paper by John J. Horton presents a compelling argument for the use of LLMs as simulated economic agents. By exploring various economic scenarios and questions, it opens up a new horizon in economics, allowing researchers to pilot studies and gain insights through simulation.

The concept of "homo silicus" offers a fresh perspective on how AI can be leveraged to understand human behavior in economic contexts. The potential applications are vast, and the implications for both economics and AI are profound.

The golden path towards superintelligence is filled with opportunities and challenges. This paper is a shining example of how AI can be harnessed to explore complex human phenomena, offering a glimpse into the future of economics and social science. 🌟

Paper Reference: "Large Language Models as Simulated Economic Agents: What Can We Learn from Homo Silicus?" by John J. Horton, MIT & NBER, January 19, 2023, arXiv:2301.07543v1