LLM Cooperation in Public Goods Games

ISEF Category: Behavioral and Social Sciences

Subcategory: Other  ·  Difficulty: Intermediate  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

People often change how they play when they think a partner is human. That reaction can shape trust, fairness, and group effort. In a public-goods game, one hidden detail can change the whole pattern of cooperation. Your project asks what happens when that partner is an LLM.

What Is It?

A public-goods game is a simple model of group teamwork. Each player gets a small amount of money or points, then chooses how much to put into a shared pool. The pool gets multiplied and shared back out. If everyone gives, the group does well. If one player free-rides, that player can gain while others lose out.

Your phenomenon asks a sharper question. What changes when one player is an LLM, a large language model that can generate human-like choices or messages? Think of it like a group project where one teammate might be a bot. If people know the partner is AI, they may act more carefully, more guardedly, or less generously. If the AI is hidden, they may treat it like a person and change course later if they suspect it.

That makes this topic a mix of game theory and behavior. Game theory studies strategic choices, while behavioral science studies what people actually do, not just what theory predicts. You can measure how cooperation shifts across rounds, across disclosure conditions, and across different message styles from the LLM.

Why This Is a Good Topic

This is a strong science fair topic because you can turn a big idea, trust in AI, into a clear experiment with real data. You can test one variable at a time, such as whether the LLM is disclosed, how cooperative it appears, or whether repeated rounds weaken generosity. The project connects to real problems in online teamwork, moderation, negotiation, and AI-mediated decision making. You can learn experimental design, survey logic, and basic stats without needing a wet lab.

Research Questions

• How does disclosure that one partner is an LLM affect average cooperation across repeated public-goods rounds?
• What is the effect of hiding the LLM identity on how quickly cooperation decays over time?
• Does the size of the trust gap change when players learn the partner is an LLM after the first round?
• To what extent do different LLM message styles change contribution levels in the next round?
• Which player traits, such as self-reported AI familiarity or trust in technology, predict stronger cooperation decay?
• What is the effect of LLM disclosure on the rate of free-riding after a cooperative first round?

Basic Materials

• Laptop or desktop computer with internet access.
• Online survey or experiment platform such as Google Forms, Qualtrics, or PsyToolkit.
• Spreadsheet software such as Google Sheets or Excel.
• Participant consent form and debrief text.
• Timer or built-in platform timing controls.
• Randomization plan for assigning conditions.
• Basic calculator for sanity checks on totals and payoffs.

Advanced Materials

• Computer with R or Python for analysis.
• Experimental platform with random assignment and branching logic.
• LLM access through an API or controlled chat interface.
• Codebook for outcome variables and condition labels.
• Statistics package for mixed-effects models or repeated-measures analysis.
• Secure data storage for anonymized participant responses.
• Pre-registration template for the study design.

Software & Tools

• Google Forms: Collects responses and assigns participants to simple conditions.
• Qualtrics: Builds randomized survey flow and repeated decision rounds.
• PsyToolkit: Runs behavioral experiments with decision tasks and timing control.
• Google Sheets: Organizes payoff data and checks contribution patterns.
• R: Fits repeated-measures models and tests whether cooperation changes across rounds.

Experiment Steps

1. Define the exact decision rule for the public-goods game and the outcome you will measure.
2. Choose the one factor you will change first, such as disclosed versus hidden LLM identity.
3. Design controls that keep the payoff structure, instructions, and round order consistent.
4. Plan how you will randomize participants and separate conditions cleanly.
5. Build a scoring system so cooperation, free-riding, and decay can be compared across rounds.
6. Decide in advance which statistical test will answer your main research question.

Common Pitfalls

• Letting the LLM change tone across participants, which makes the manipulation mix identity with personality.
• Changing payoff rules between rounds, which breaks the repeated-game comparison.
• Telling participants too much about the hypothesis, which can push them to guess the desired behavior.
• Using too few rounds, which makes cooperation decay hard to measure.
• Failing to separate the hidden and disclosed conditions cleanly, which leaves you unsure what caused the behavior change.

What Makes This Competitive

A stronger project will do more than compare two averages. It will measure how cooperation changes round by round, test whether the effect survives controls, and check whether the LLM disclosure changes trust, fairness, or free-riding in different ways. You can make it more competitive by using a preregistered design, a clear manipulation check, and a model that handles repeated choices instead of treating every response as independent. A novel twist, like comparing neutral versus highly human-like LLM messages, can add depth without making the study messy.

Project Variations

• Test whether the effect changes when the LLM sends no explanation versus a short strategic explanation.
• Compare first-round cooperation in a one-shot public-goods game versus a repeated game with the same partner.
• Swap the sample type by testing students, adults, or mixed-age online participants to see whether AI familiarity changes the result.

Learn More

• PubMed: Search for review articles on public goods games, cooperation, trust, and human-AI interaction.
• MIT OpenCourseWare: Search economics and game theory course materials for notes on repeated games and strategic behavior.
• PLOS ONE: Search for open-access studies on cooperation, trust, and decision making.
• Frontiers in Psychology: Search for open-access articles on social decision making and human-AI behavior.
• arXiv: Search for preprints on LLMs, strategic interaction, and behavioral experiments.