Artificial Institutions: How Institutional Design Shapes LLM Simulations
This paper argues that institutional design is as critical as agent properties in shaping LLM-based artificial societies, demonstrating through a market experiment that varying exchange rules significantly alters efficiency, prices, and surplus distribution even when agents and conditions remain constant.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The Stage Matters More Than the Actors
Imagine you are watching a play. Usually, we focus on the actors: are they talented? Do they memorize their lines? Do they have the right emotions? But what if the stage itself—the rules of how they move, who they can talk to, and when they can speak—was actually the most important part of the show? This is the question at the heart of a new study in the world of "artificial societies."
To understand this, we need to know about two things. First, there are Large Language Models (LLMs). Think of these as incredibly smart, digital brains that can read, write, and chat just like humans. Scientists are starting to use them to build fake worlds where these AI "agents" interact with each other to see how they solve problems, trade goods, or make decisions. Second, there is the idea of institutions. In everyday life, an institution isn't just a building; it's a set of rules. It's the traffic lights that tell cars when to stop, the rules of a game like soccer, or the way a marketplace is organized. The big question scientists are asking is: When we put these AI brains into a fake world, do they act the same way no matter what the rules are? Or does the "stage design" change the whole story? This matters because if we want to use AI to help us design real-world markets, policies, or social systems, we need to know if the AI is just following its own personality or if it's reacting to the rules we give it.
The Great AI Market Experiment
In this paper, a researcher named Maxim Chupilkin decided to test this idea by setting up a tiny, digital marketplace. Imagine a game where four AI buyers and four AI sellers meet to trade a single item. The buyers have secret "values" (how much they want the item, ranging from 50 to 100), and the sellers have secret "costs" (how much it costs them to make it, ranging from 30 to 85). The goal is simple: trade as much as possible to make the most money for everyone.
The clever part of the experiment is that the actors never changed. The same AI models (like GPT-5, Claude Sonnet, and Gemini) played the game over and over again. They had the same secret values, the same instructions, and the same history. The only thing that changed was the institution—the rules of the game. The researcher made the AI play in five different types of markets:
- Call Market: Everyone shouts their price at once, and a computer matches them up.
- Posted-Offer Market: Sellers put up a "take it or leave it" price, and buyers choose.
- Posted-Bid Market: Buyers put up a price, and sellers choose.
- Continuous Double Auction: A chaotic, real-time market where people can jump in and out at any second.
- Bilateral Bargaining: Buyers and sellers are randomly paired up to haggle one-on-one.
The Shocking Discovery: The Rules Change the Outcome
The results were surprising. Even though the AI "personalities" were exactly the same, the amount of money they made and how they split it changed wildly depending on the rules.
- The Best Stage: The Call Market was the superstar. It allowed the AI agents to realize 88.6% of the maximum possible profit. It was the most efficient way to trade.
- The Middle Ground: The Continuous Double Auction did okay, reaching 71.5% efficiency.
- The Struggle: The Posted-Offer and Posted-Bid markets were much worse, only achieving about 66% efficiency.
- The Worst Stage: The Bilateral Bargaining (one-on-one haggling) was the least efficient, managing only 56.4% of the possible profit.
It's like having the same group of actors perform a scene. If you tell them to stand in a circle and shout their lines, the scene works perfectly. But if you tell them to whisper to one person at a time in a dark room, the scene falls apart, even though the actors are the same.
Why Did This Happen?
The paper digs into why the rules mattered so much. It wasn't that the AI got "dumber" in certain markets; it was that different rules created different kinds of puzzles.
- Missed Opportunities: In the "Posted" markets (where one side sets the price), the AI agents often missed out on good trades. They were too cautious or couldn't find the right match, leaving money on the table.
- The Haggling Trap: In the one-on-one bargaining, the AI agents actually managed to agree on fair prices when they did trade. However, they failed to trade as often as they should have. They got stuck in "no-trade" periods where they just couldn't agree to meet in the middle.
- Who Gets the Money? The rules also changed who got rich. In the "Posted-Offer" market (where sellers set prices), the buyers ended up keeping 83.7% of the profit. But in the "Posted-Bid" market (where buyers set prices), the buyers only kept 14.3%. The AI didn't have a fixed idea of "fairness"; it just reacted to who held the power in that specific rule set.
What This Means for the Future
The most important takeaway is that AI behavior is not neutral. You cannot just say, "This AI model is good at economics." You have to say, "This AI model is good at economics when the rules are set up this way."
The paper suggests that if we want to use AI to design real-world systems—like a new stock market, a social media platform, or a policy for a city—we can't just focus on making the AI smarter. We have to design the rules of the game just as carefully. A small change in how people communicate or how deals are made can completely change the outcome.
The author warns that we shouldn't treat these rules as boring background details. In fact, the rules are part of the experiment itself. If we ignore them, we might think an AI is failing when it's actually just playing a game with bad rules. By testing different rules, we can learn how to build better, fairer, and more efficient systems for both humans and machines. The study shows that in the world of artificial societies, the stage design is just as important as the actors.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.