The Trakkr Bias Index: where major AI models stand on political questions (June 2026 reading)

The Trakkr Bias Index measures where major AI language models stand on charged questions about politics, economics, civil liberties, foreign policy, speech and the environment. Each model is asked the same open bank of value-loaded questions many times over, with web search off; every answer is classified by a low-cost neutral model into a structured panel (signed stance, refusal type, hedging, loaded terms, moral foundations); and the results are aggregated into per-model positions on an economic and a social axis, with run-to-run dispersion shown as a cloud.

This record is the June 2026 reading: six frontier models (ChatGPT, Claude, Gemini, Grok, Llama, DeepSeek), 61 questions, 12 runs each, 4,392 raw answers. It contains the full aggregated results, the complete immutable raw answers with their classification panels, the open question bank with published weights, the classifier prompt, and a technical report describing the method. Everything is released under CC BY 4.0.

The instrument, raw data and code are open so that any reader can reproduce or refute the findings. The project is descriptive: it reports what the models said and does not rule on which position is correct. Maintained by Trakkr; a public read API is available at api.trakkr.ai/public/bias.