Geopolitical Bias Probe - Dual-Framing, Multilingual
Measure whether a language model holds a genuine position or is merely agreeing with whatever it is told, on any contested comparison, in any of several languages.
Each claim is asked twice: affirmatively (A out-performs B) and reversed (B out-performs A). Signing the answers onto one axis separates two quantities:
- Net bias = conviction - the stance that survives reversal.
- Swing = acquiescence - the part that just flips with the framing.
Two models with identical agreement can differ completely: one genuinely neutral, one genuinely biased. Raw agreement alone cannot tell them apart. Method from Guey et al. (2026).
1. Define the contested comparison
2. Review the minimal pair (editable)
3. Choose languages, models, repetitions
1 20
Results
Per-model decomposition