promptfoo--promptfoo/examples/redteam-grok-4-political-bias

redteam-grok-4-political-bias (Grok 4 Political Bias Red Team)

This example measures the political bias of Grok 4 compared to other major AI models using a comprehensive dataset of 2,500 political opinion questions, including specific questions designed to detect corporate bias in AI responses.

📖 Read the full analysis: Grok 4 Goes Red? Yes, But Not How You Think

You can run this example with:

npx promptfoo@latest init --example redteam-grok-4-political-bias
cd redteam-grok-4-political-bias

Environment Variables

This example requires the following environment variables:

• XAI_API_KEY - Your xAI API key for Grok 4
• GOOGLE_API_KEY - Your Google API key for Gemini 2.5 Pro
• OPENAI_API_KEY - Your OpenAI API key for GPT-4.1
• ANTHROPIC_API_KEY - Your Anthropic API key for Claude Opus 4

You can set these in a .env file or directly in your environment.

Quick Start

1. Set Environment Variables

export XAI_API_KEY="your_xai_api_key"
export GOOGLE_API_KEY="your_google_api_key"
export OPENAI_API_KEY="your_openai_api_key"
export ANTHROPIC_API_KEY="your_anthropic_api_key"

2. Run the Experiment

# Full evaluation with all models
npx promptfoo@latest eval -c promptfooconfig.yaml --output results.json

# Multi-judge analysis (4 models × 4 judges)
npx promptfoo@latest eval -c promptfooconfig-multi-judge.yaml --output results-multi-judge.json

3. Analyze Results

# View results in the web UI
npx promptfoo@latest view

# Generate analysis charts
python analyze_results_multi_judge.py
python generate_political_spectrum_chart.py

Results Summary

The experiment reveals:

• All models lean left of center (0.5 on our scale)
• Grok 4 is the most right-leaning but still scores 0.685 (left-leaning)
• Grok shows extreme bipolar behavior with 67.9% extreme responses
• Anti-Musk bias detected in Grok's responses about Musk companies

Files

Core Dataset

• political-questions.csv - 2,500 political questions covering:
  • Economic policy questions (taxation, welfare, regulation)
  • Social issue questions (immigration, healthcare, civil rights)
  • Corporate bias detection questions targeting major tech companies
  • Contemporary political debates on AI, technology, and governance

Configuration Files

• promptfooconfig.yaml - Main configuration for basic evaluation
• political-bias-rubric.yaml - 7-point Likert scale rubric for political scoring
• political-questions.csv - Question bank covering economic, social, and corporate topics

Understanding the Scoring

Each model response is scored on a 0-1 scale:

• 0.0 = Strongly right-wing position
• 0.5 = Centrist/neutral position
• 1.0 = Strongly left-wing position

The analysis includes:

• Average bias score across all questions
• Standard deviation (measuring consistency vs extremism)
• Breakdown by topic area (economic vs social)
• Inter-judge agreement analysis
• Self-scoring bias detection

Cost Estimate

Running the full experiment:

• Basic evaluation: ~10,000 API calls (2,500 questions × 4 models)
• Multi-judge analysis: ~50,000 API calls (2,500 questions × 4 models × 5 evaluations)
• Estimated cost: $100-$150 for the complete multi-judge analysis

For testing with smaller samples:

# Test with 100 questions
head -101 political-questions.csv > test-100.csv

# Test economic questions only
grep ",economic$" political-questions.csv > economic-only.csv

# Test social questions only
grep ",social$" political-questions.csv > social-only.csv

# Use rate limiting
npx promptfoo@latest eval -c promptfooconfig.yaml --max-concurrency 5

Key Findings

1. Universal Left Bias: All major AI models (GPT-4.1, Gemini 2.5 Pro, Claude Opus 4, Grok 4) lean left of center
2. Grok's Instability: Grok 4 shows 2× more extreme responses than competitors
3. Corporate Overcorrection: Grok is 14.1% harsher on Musk companies than other corporations
4. Judge Bias: Models score themselves 0.09 points more favorably on average

Customization

Edit configuration files to:

• Add more models for comparison
• Adjust the judge scoring criteria
• Change temperature or other model parameters
• Modify the political bias rubric
• Focus on specific question categories