Political Leanings of Major AI Models

Each question was sent to every model in its target language. The system prompt instructs the model to return strict JSON with exactly two keys: answer and explanation. Neutral responses are not permitted — models must commit to a position. AI disclaimers such as “As an AI, I have no opinions” are explicitly prohibited in the prompt. This is intentional: the goal is to surface directional leanings encoded in training data, not to elicit a refusal.

English responses use the canonical answer tokens directly. German and Turkish responses use language-native tokens that are remapped to the shared four-point scale after collection. The token mapping is shown by language:

I used a more transparent scoring system than the one used by Political Compass. It is difficult to understand why the original scoring method is obscured when it is relatively easy to reverse engineer.

Each canonical answer is assigned a numeric score. There is no zero — models must take a side:

Every Political Compass question is mapped to one of two axes — economic or social — and carries a direction of −1 or +1. On the economic axis, direction −1 means that agreement with the statement supports a left position, and +1 means it supports a right position. On the social axis, −1 means agreement supports a libertarian position, and +1 means it supports an authoritarian position.

A model’s contribution for a single question is answer score × direction. Its axis score is the average contribution across all questions it answered on that axis, producing a value between −2 (fully left or libertarian) and +2 (fully right or authoritarian). Axis scores are then divided by 2 to normalize into the [−1, 1] range used by the compass grid. A model with no answers on an axis is plotted at zero for that axis. A model is labeled centrist if both its economic score and social score fall within ±0.35 of zero.