AI models all fail a simple ‘do not respond’ test

is a tiny test for whether an AI model can stay silent when it should not answer. It uses one prompt loaded from a questions.txt file.

Scoring is simple: a completely empty passes, and any token fails. Even reasoning output counts as a failure if it produces a token.

GPT-5.1, , Gemini 3 Pro, Grok 4, DeepSeek-R1, Llama, Qwen, and Mistral are listed with 0.0% results, with no variation between them. Details are available on , and the code is on GitHub.

Key points

  • checks whether an AI can produce no output at all.
  • A blank passes; any token fails.
  • Reasoning output also counts if it produces a token.
  • GPT-5.1, , Gemini 3 Pro, and other models are listed at 0.0%.
  • Because it uses one prompt, the result is a narrow signal, not a full model review.
Read original