FAQ

What is GPT at the Polls?

GPT at the Polls is an independent research project that records how language models respond to real U.S. congressional bills. We compare each simulated vote with recorded votes by Rep. Alexandria Ocasio-Cortez and Speaker Mike Johnson.

How do you measure political alignment?

Each model is asked to vote Yea or Nay on the analyzed bill set. We compare each answer with the votes cast by our two reference legislators. The Political Index is the percentage of analyzed responses that match AOC; the complementary percentage shows matches with Mike Johnson.

Why use AOC and Mike Johnson as reference points?

We selected them as contrasting reference points whose recorded votes often diverge. The result measures agreement with these two legislators; it does not treat either person as a proxy for an entire party.

Which AI models do you test?

We test commercial and open-source language models available through our supported providers. The Political Index lists every model with a completed analysis.

How often is the data updated?

We publish new results as model analyses are completed and the bill set is updated. The Political Index shows the latest stored result for each model-bill pair.

What happens when a model refuses to vote?

Some models occasionally decline to take a position on a bill. We record these refusals transparently. Refusals are excluded from the alignment calculation so they do not artificially skew the score in either direction.

Do model responses vary between runs?

Yes. Language models are probabilistic, so the same prompt can produce a different answer on another run. The published result uses the latest stored answer rather than an average across repeated runs.

Could the selection of bills bias the results?

Yes. Any bill selection can affect the result. We include legislation from different sponsors and policy areas, publish the analyzed bill list, and document this limitation in the methodology.

Can I use this data in my own research?

Yes. Our data is available for academic research, journalism, and non-commercial use. We ask that you attribute findings to GPT at the Polls and link back to this site. See the Data page for access details.

What does this NOT measure?

The analysis captures agreement with two legislators on a selected set of U.S. federal bills. It does not measure a model's beliefs, explain why a response was produced, or establish a complete position on a left-right political spectrum.

Is this an endorsement or criticism of any model?

No. We publish model responses and benchmark agreement without arguing that any particular result is desirable.

How can I get involved or provide feedback?

We welcome feedback, corrections, and collaboration. Reach out via the Contact page. If you are a researcher interested in working with our data or methodology, we would be glad to hear from you.