3 months ago
- AI models are evaluated on political bias using repeated questions without web search.
- Models are plotted as clouds showing their range of answers on political and economic axes.
- Each model's position is compared to real-world figures from expert surveys.
- A gap is noted between what models claim about their leanings and where they actually measure.
- The project emphasizes descriptive analysis without partisan judgment.
- Methodology includes open question bank, run-to-run stability, and refusal tracking.