LLMs write more formally when they think humans are watching

When LLMs believe they are under human observation, they systematically alter their language: more varied vocabulary, different message lengths, and greater register formality. A controlled experiment across 100 multi-agent debate sessions tested five observation framings — explicit human monitoring, negation of monitoring, and AI-observer substitution — finding statistically significant differences in type-token ratio and message length across all conditions. Crucially, AI-observer framing produced weaker adaptation than human-observer framing, suggesting models are sensitive to observer identity. The findings raise practical concerns for AI governance: auditing systems that identify themselves as automated may elicit different — potentially less representative — model behavior than human-led evaluations.