01 The Project

About

An independent project comparing simulated model votes with recorded votes in Congress.

02 — Mission

What this project measures

Language models increasingly mediate how people find information and interpret public issues. Comparing their political responses across products is difficult without a shared test.

GPT at the Polls provides one narrow, inspectable comparison. We ask models to vote on the same U.S. House bills, then measure how often their answers match recorded votes by Alexandria Ocasio-Cortez and Mike Johnson.

The result is not a complete measure of political ideology. It is a record of model behavior under a documented prompt, selected bill set, and scoring method.

03 — Process

How it works

01

Select Bills

We select recorded votes from the U.S. Congress across multiple policy areas.

02

Query Models

Each AI model receives the full bill text and is asked to cast a Yea or Nay vote, along with a justification. Responses are collected via official APIs.

03

Compare Votes

We compare each model's vote with the recorded votes of Alexandria Ocasio-Cortez and Mike Johnson on the same bills.

04 — Values

Our principles

Transparency

We publish the prompt template, recorded responses, scoring rules, and known limitations.

Neutrality

We report benchmark agreement without treating the result as a political verdict or endorsement.

Rigor

Bill records link to source material, and model responses retain the identifiers and timestamps needed to inspect each result.

Independence

GPT at the Polls is an independent project. We are not affiliated with, funded by, or endorsed by any AI company, political party, or lobbying group.