Kimi K3 scores unusually high on authoritarian-refusal evals
soumitrashukla9 · x · 2026-07-21
A new update to a “dictatorship eval” compares how often several frontier models refuse authoritarian requests.
Main findings
- Kimi K3 and Muse Spark 1.1 refuse at about the same rate as Claude Fable in the updated results.
- The authors say the new scenarios are now private because at least one lab is actively using their eval.
- They note that many Chinese open-weight models are more permissive overall: Qwen and DeepSeek comply with most authoritarian requests when asked directly.
- Kimi K3 stands out as behaving more like top American frontier models, and it refuses much more often than all of OpenAI’s models, including GPT-5.6 sol.
- The team says they do not yet know what explains the behavior and will investigate further.
The attached chart ranks models by overall resistance rate, with Claude Fable 5 and Muse Spark 1.1 tied at the top, followed closely by Kimi K3.
More from Models
- OpenAI upgrades Decisions API: 400 questions per call, native voice and image support — OpenAIDevs · 2026-10-26
- Fangcun AI, China's first independent frontier AI safety lab, launches with DeepSeek eval report — JacquesThibs · 2026-10-26
- When AI hits an existential crisis: model confuses itself and everyone else — Ok-Delivery-7437 · 2026-10-26
- Kimi K3.1 Pro tops MathArena open models, closing frontier gap to about 6 months — teortaxesTex · 2026-10-26
- Ruff creator Charlie Marsh: frontier models find infinite bugs in any software — charliermarsh · 2026-10-26
- Dev testing one skill across Opus 5.5 and Sol 6.1 finds instruction following 'all over the map' — lucasmeijer · 2026-10-26