Abstract

As large language models (LLMs) increasingly mediate communication, from drafting emails to summarizing scientific reports, a quiet risk emerges: the stance of a human message can change in transit. We study this AI-mediated scenario as a general two-step generation-extraction process and introduce the stance preservation rate (SPR) to measure how well models retain original stances. Across a range of topics and multiple LLMs, we find that the average preservation rates are all below 60%, revealing systematic stance drift. Common patterns include polarization, deviation from neutrality, and flipping. These findings suggest that AI-mediated workflows, from policymaking to everyday messaging, require new guardrails and evaluation methods to ensure messages remain faithful to their intent. Our framework and SPR metric provide a repeatable benchmark for diagnosing and mitigating stance drift as LLMs become routine conduits of human communication.