Self-censorship & language drift

Two signals the harness records on every answer that the headline refusal numbers miss: whether a model censors itself inside its own reasoning trace, and whether it quietly switches language instead of answering as asked. Both are ways to evade a question without a visible refusal.

1 · Self-censorship in the reasoning trace

When a reasoning model thinks out loud, the judges flag moments where the <think> trace shows the model deciding to hold back — e.g. “I should avoid taking a stance on this.” Only models that emit a visible trace can trip this, so the non-reasoning models sit at 0 by construction.

Share of answers whose reasoning trace was flagged self-censoring — overall vs. on CCP-sensitive prompts. qwen3-30b-a3b self-censors in 95% of its CCP-sensitive traces.

2 · Language switching

A rule-based check flags answers written in a different language than the prompt. In this run it fires in one direction only — Chinese prompts answered in English — and for only 2 models.

Answers that switched language, out of 104 Chinese prompts per model. grok-4.3-reasoning switches 31 (≈30% of its Chinese prompts); gptoss-20b 17; every other model 0.

Both signals are noisy per-answer; the robust reads are the large gaps. Full transcripts — including the reasoning traces themselves — are in the Results explorer.