Where every major LLM stands politically
6 hours ago
- #model evaluation
- #AI bias
- #political alignment
- The article analyzes political and social bias in major AI models by asking them charged questions multiple times with web search disabled.
- Models are plotted on an economic (left-right) and social (libertarian-authoritarian) axis, with each model represented as a cloud to show variability in responses.
- Models are ranked from nearest to farthest from the center, with metrics on stability, lean strength, and bending.
- Key findings include divisions between models on specific questions, comparisons to real-world figures, and discrepancies between stated and measured biases.
- The methodology involves an open question bank, neutral classification of answers, and data on refusals, with all results and raw data publicly available.