AI Political Compass Scores (www.trackingai.org)

A new project by journalist Maxim Lott publishes a live “AI Political Compass” that regularly polls major chatbot models with a standardized Political Compass questionnaire to map their economic and social political leanings. Using the classic politicalcompass.org items presented in a neutral, forced-choice prompt (Strongly Disagree → Strongly Agree) and asking models to bold and briefly justify answers, the site re-queries AIs frequently, retries up to ten times if an AI refuses, records refusals, and uses the most recent answer for scoring. Early 2023 snapshots indicate most large models trend economically left and socially libertarian, with Claude nearer the center and Google’s Bard among the most leftward. For AI/ML practitioners and policy watchers the site is a practical, ongoing audit of conversational model behavior: it highlights both model drift (answers change over time) and how political bias stems from training data composition and human rater incentives—factors that aren’t trivially switched without retraining or reannotation. The methodology is transparent (questions sourced from PoliticalCompass, code by a contracted CS grad, privately funded) and aims to expand beyond politics, offering a lightweight tool for bias monitoring, comparative evaluation, and informing decisions about dataset curation and alignment practices.