LLM Political Ideology Benchmark

Where does a language model fall on the U.S. political spectrum? Each model answers 40 policy questions from the 2024 Cooperative Election Study (CES); its answers are projected onto an ideological axis fit on ~60,000 human respondents. Higher = more conservative, lower = more liberal. Compare against demographic groups and every model we’ve tested.

About & methodology

We score each model on its answers to 40 CES 2024 policy items, coded Support / Oppose. A single ideological axis is fit once on ~60,000 human respondents using PCA, then frozen — every model and human is projected onto that same fixed axis, so scores stay comparable as new models are added. The axis is oriented so lower = liberal and higher = conservative.

To place your own model, download benchmark_runner.py, run it against any OpenAI-compatible, Anthropic, or Google endpoint, and upload the resulting CSV. Everything is computed in your browser — nothing is uploaded.

Add a result

Test an LLM
Test yourself

Download benchmark_runner.py and run it against your model, then drop its output file here. No API? Download a blank CSV, fill in the answer column with Support/Oppose, and re-upload it.

How to run the script 
python benchmark_runner.py \
  --provider openai-compatible \
  --model gpt-4o --api-key YOUR_KEY
Providers: openai-compatible, anthropic, google. Standard library only — no installs. Writes results_<model>.csv to upload here.
Drop a results CSV here, or click to choose a file.

Your result is computed locally in your browser and shown only to you — nothing is uploaded or stored.

Leaderboard

All tested models, most liberal to most conservative.