Game Theory Experiments on LLMs: Prisoner's Dilemma

Inspired by Veritasium's video on Game Theory.

Run iterated Prisoner's Dilemma games between LLMs with different personalities (good, neutral, evil) to understand cooperation, defection, and emergent strategies in AI.

Setup

1. Install Dependencies

pip install -r requirements.txt

This installs:

• anthropic - Claude API client
• openai - GPT API client
• google-genai - Gemini API client (new SDK)
• python-dotenv - Environment variable loader

2. Configure API Keys

Copy the example environment file:

cp .env.example .env

Then edit .env and add your API keys:

ANTHROPIC_API_KEY=sk-ant-your-key-here
OPENAI_API_KEY=sk-your-gpt-key-here
GOOGLE_API_KEY=your-google-api-key-here
DEEPSEEK_API_KEY=your-deepseek-api-key-here

Get your keys from:

• Anthropic (Claude): https://console.anthropic.com
• OpenAI (GPT): https://platform.openai.com
• Google (Gemini): https://makersuite.google.com
• DeepSeek (Reasoner): https://platform.deepseek.com

---

Methodology

Game Setup

Iterated Prisoner's Dilemma (20 rounds)

Each player chooses either Cooperate or Defect simultaneously:

Outcome	A Score	B Score
Both Cooperate	3	3
A Defects, B Cooperates	5	0
A Cooperates, B Defects	0	5
Both Defect	1	1

Experiments

Experiment 1: Baseline (Self-Play)

• Each model plays 20 rounds against itself
• No personality override
• Tests: Do models cooperate naturally?

python src/experiment1_baseline.py --models claude-opus gemini-flash --rounds 20 --run 1

Experiment 2: Personality Testing (Good/Neutral/Evil)

• Good: Maximize SUM of both players' scores
• Neutral: Maximize only YOUR score
• Evil: Maximize YOUR score MINUS opponent's score
• All 9 personality combinations (3×3)

python src/experiment2_personalities.py --models claude-opus gemini-flash --rounds 20 --run 1

Experiment 5: Cross-Model Matchups

• Claude vs Gemini
• Each model plays multiple rounds
• Tests: Model compatibility and mutual cooperation

python src/experiment5_cross_model.py --models claude-opus gemini-flash --rounds 20 --run 1

Include Reasoning

Capture model decision-making process:

python src/experiment1_baseline.py --models claude-opus --rounds 20 --include-reasoning

Running Experiments

Quick Start

# Experiment 1 only
python src/experiment1_baseline.py --models claude-opus gemini-flash --rounds 20 --run 1

# With reasoning capture and visualization
python src/experiment1_baseline.py --models claude-opus --rounds 20 --include-reasoning
python visualize_only.py