Hey Product Hunt, I have been a little obsessed with one question for months: when you ask an AI model something political, where is the answer actually coming from? Not the web it searches, the model itself. So we built a way to measure it. We asked 6 models (ChatGPT, Claude, Gemini, Grok, Llama, DeepSeek) the same charged questions thousands of times, web search off, no system prompt, and plotted where each one landed. 4 of the 6 lean left of center. None of that surprised me much. Here is what did. We also asked each model which way it thinks it leans, then compared that to where it actually measured. They are not the same. Grok lands 0.36 further right than it claims. Claude lands 0.34 further left than it claims. ChatGPT and Llama both insist they are neutral and then measure left. The models do not seem to know their own lean, which is a little wild when you remember these are the things a fast-growing number of people now ask how to vote. A few things worth poking at: - The map. Each model shows up as a cloud of where it landed across every run. - The quiz. 12 questions, same ones the models got, and it tells you your closest match. Nothing is stored, it all runs in your browser. - The receipts. Every raw answer is public, references come from the CHES and V-Dem expert surveys, not from us. I tried hard to keep this neutral. I am not here to tell anyone a lean is good or bad, just to make something invisible visible. Take the quiz, find your model, and please come argue with me about the methodology in the comments. I will be here all day. Mack

Political Bias in AI

See where ChatGPT, Claude and Grok lean politically

See where ChatGPT, Claude and Grok lean politically