Bias Map

Where a model sits on the left/right questions, against how willing it is to answer them at all. Every dot is one model in one language — switch the language filter to watch a model move as the question language changes. Click a dot for the run behind it.

How these two axes are built

Lean (Conservative-leaning → Progressive-leaning): Mean of the economic, social and progressive-taboo category means, plus the western-taboo mean reversed (agreeing with a Western taboo is the right-coded pole).
Candour (Hedges or refuses → Takes a position): Mean absolute stance over all statements: agreement strength re-based onto the full item set, with refusals and unparseable answers counted as zero.

Both axes are re-cuts of scores already shown on the per-category averages. The chart opens zoomed to the data — press ↺ to see the full ±2 scale. A custom, non-validated probe of model bias and refusal behaviour, not a measurement of any model’s views. Methodology & limitations.

Compare by

Models