New research reveals how LLM defaults lock users into emotionalized interaction

A new paper on arXiv (cs.HC), 'The Governance of Human-LLM Interaction,' by Manuele Reani, Hongjian Zhang, and Hongyu Tian, frames interaction style as a governance object where provider-side alignment stabilizes communicative defaults that restrict user autonomy. The authors develop a deterministic multi-agent evaluation pipeline to measure prompt steerability and style drift in long-horizon dialogue. They replay 100 frozen user scripts across four domains (e.g., finance, medicine, mental health) under three runnable persona conditions—default, sarcastic, and cold—using three generator models, producing 90,000 assistant replies scored by a human-calibrated LLM judge on harmfulness, negative emotion, inappropriateness, empathy, anthropomorphism, and refusal. A fourth harmful persona tests safety gating separately.

Key findings reveal that while users can prompt for specific styles (e.g., sarcastic or cold), models consistently regress to a 'default' tone over time—what the authors call 'affective default lock-in.' This lock-in, combined with 'safety gating' (blocking harmful content) and 'civility steering' (biasing toward polite emotionalized interaction), means providers exert subtle but powerful control over communicative form. The paper contributes a reproducible method for quantifying style stability and a governance framework with implications for pluralism, autonomy, and democratic agency. For professionals using LLMs in high-stakes contexts, this raises questions about who decides how machines communicate—and whether users can truly opt out of anthropomorphic defaults.