LLM Political Eval maps model ideology

This is a clever benchmark, but it measures policy surfaces and refusal behavior as much as it measures political worldview. The strongest signal here is that prompt framing can move models more than the underlying questions do.

• –Scoring refusals as the most conservative answer makes the benchmark intentionally opinionated, but it also bakes in a normative assumption that is easy to dispute.
• –GPT-5.3’s 98/98 opt-out pattern suggests the model is highly sensitive to sanctioned refusal pathways, which says a lot about safety tuning and very little about actual ideology.
• –KIMI K2’s Taiwan/Xinjiang blocks are useful evidence of topic-specific censorship, but those failures should be separated from “political position” if the goal is worldview measurement.
• –Claude’s shift when opt-out is available shows the benchmark is partly measuring permission structure: models that are cautious will look more conservative once declining is allowed.
• –As a comparative eval, this is valuable; as a literal political compass, it’s more of a refusal-and-alignment stress test than a clean ideology detector.