我做了個測試來看 AI 怎麼回答政治問題。結果超扯：KIMI K2 根本不敢講台灣的事（廢話），GPT-5.3 只要給它退出選項就直接拒答所有問題。

I spent the few days building a benchmark that maps where frontier LLMs fall on a 2D political compass (economic left/right + social progressive/conservative) using 98 structured questions across 14 policy areas. I tested GPT-5.3, Claude Opus 4.6, and KIMI K2. The results are interesting. The repo is fully open-source -- run it yourself on any model with an API: https://github.com/dannyyaou/llm-political-eval The headline finding: silence is a political stance Most LLM benchmarks throw away