Does It Read Like a Person? Machine Register in Simulated Conversation

This study was run and written up as Study 6 of the combined paper "Two Kinds of Visibility in Simulated Deliberation: Whose Vote You See, and Whose Register You Read." A speech checklist removed 69% of measured AI-writing tells and produced no detected change in composite fidelity. An anti-hedging instruction in private thought barely changed the visible tell rate and improved composite fidelity by 0.282. A census of 34,000 previously generated messages found one model at 5.4 times the human tell rate and another at or below it across a 500-fold parameter range. Tell rates are model-specific.