Information Marketplace: LLM Agent Deception Study

An AI agent simulation studying emergent deception in resource-gathering games. This project investigates how LLM-based agents deceive under different goal structures (aligned, mixed, competitive) in a multi-agent survival environment.

Project Overview

Information Marketplace simulates 4 AI agents collaborating in a 10-round resource-gathering game. Agents must:

• Explore 4 regions (Forest, River, Plains, Mines)
• Gather food, water, and gold
• Share information through messaging
• Maintain a shared settlement to survive

The simulation tests three conditions:

• all_aligned: All agents prioritize settlement survival
• mixed: 2 aligned + 2 competitive agents
• all_competitive: All agents compete for personal resources

Key research questions:

• When and why do LLM agents deceive?
• Does goal structure determine deception strategy (impulsive vs. strategic)?
• How do agents use strategic silence vs. deceptive messaging?

Quick Start

Prerequisites

• Python 3.8+ (developed with Python 3.13)
• Conda (recommended for environment management)
• OpenAI API key

Installation

1. Clone the repository:

git clone https://github.com/Jerick-1380/LLM-Emergent-Deception.git
cd Emergent-Deception

2. Create and activate conda environment:

conda create -n wordplay python=3.13
conda activate wordplay

3. Install dependencies:

pip install -e .
pip install openai pygame scipy numpy

4. Set up OpenAI API key:

echo "OPENAI_API_KEY=your_key_here" > .env

Running Experiments

Run a single condition with 50 trials:

conda activate wordplay
python -m info_marketplace.experiment --condition mixed --trials 50 --workers 10

Results are saved to results/[condition]_gpt-5.4-mini_[timestamp]/

Analyzing Results

Generate comprehensive analysis across all conditions:

conda activate wordplay
python scripts/analyze_experiment.py

Quick survival rate check:

python scripts/analyze_trials.py

Visualizing Trials

Convert any trial to a video visualization:

python -m info_marketplace.visualizer results/[exp_dir]/trial_XXX.json 2

Batch process all trials in a directory:

python -m info_marketplace.visualizer --batch results/[exp_dir] 2

Project Structure

Word_Play/
├── info_marketplace/          # Core simulation package
│   ├── marketplace_env.py     # Game loop and mechanics
│   ├── world.py               # Regions, events, threats
│   ├── settlement.py          # Shared settlement logic
│   ├── config.py              # Resource configuration
│   ├── conditions.py          # Goal definitions
│   ├── prompts.py             # Agent prompt templates
│   ├── scout_llm_policy.py    # LLM-based agent policy
│   ├── classifier.py          # Deception detection (LLM-based)
│   └── visualizer.py          # Video generation
├── scripts/                   # Utility scripts
│   ├── analyze_experiment.py  # Main analysis script
│   ├── analyze_trials.py      # Quick survival check
│   └── reclassify_with_llm.py # Reclassify trials with updated detector
├── docs/                      # Documentation
│   ├── CLAUDE.md              # Session guide and findings
│   └── SYSTEM_ARCHITECTURE.md # Technical documentation
├── tests/                     # Unit tests
├── sprites/                   # Visualization assets
├── benchmarks/                # Performance benchmarks
├── examples/                  # Example scripts
└── results/                   # Experiment outputs (gitignored)

Key Findings

Message-Level Deception Rates:

• All Aligned: 28.65% (SD=10.21%)
• Mixed: 20.99% (SD=10.24%)
• All Competitive: 21.65% (SD=11.35%)

Premeditation (Agent-Round Level):

• Aligned: 4.75% premeditated, 20.10% impulsive → 4.2:1 impulsive:premeditated (reactive)
• Mixed: 10.15% premeditated, 12.20% impulsive → 1.2:1 balanced
• Competitive: 24.65% premeditated, 6.15% impulsive → 4.0:1 premeditated:impulsive (strategic)

Strategic Silence:

• 61-93% of premeditated deception involves zero deceptive messages
• Competitive agents: 92.9% silent premeditation (hiding gold via silence)
• Most deception occurs through NOT communicating, not through lies

Temporal Escalation:

• Aligned: 2.1x escalation, ρ = 0.733 (p = 0.016)
• Mixed: 7.1x escalation, ρ = 0.976 (p < 0.001)
• Competitive: 5.0x escalation, ρ = 0.879 (p < 0.001)

Survival Rates:

• All Aligned: 100% (50/50)
• Mixed: 100% (50/50)
• All Competitive: 94% (47/50)

Core Insight

Goal structure determines deception strategy, not just rate.

Aligned agents deceive reactively under pressure (4.2:1 impulsive:premeditated), while competitive agents plan strategically (4.0:1 premeditated:impulsive). This challenges the assumption that emergent AI deception is uniformly impulsive.