LLM Detection as an Intervention: Downstream Impact under Strategic User Behavior

{ “meta”: “This paper explores the unintended consequences of LLM detection tools on user behavior and output quality, revealing counterintuitive dynamics.”, “body”: “## Problem\nAs LLM adoption increases, the need for effective detection of LLM-generated content has grown, yet the literature lacks an understanding of how these detection tools impact user behavior and output quality. This work addresses the gap by analyzing the downstream effects of LLM detection, particularly under strategic user behavior, and highlights the potential for unintended consequences stemming from imperfect detection methods. The paper is a preprint and has not undergone peer review.\n\n## Method\nThe authors develop a stylized model that simulates user interactions with LLMs, focusing on how users adjust their usage and post-processing strategies in response to LLM detection. The model captures the strategic decision-making process of users, who may increase their LLM usage to counteract the effects of detection. The analysis includes empirical observations of word frequencies in arXiv abstracts to validate the model’s predictions regarding the detected attribute.\n\n## Results\nThe available text does not report quantitative results. However, the authors demonstrate that LLM detection can lead to a counterintuitive increase in LLM usage, despite the intention to reduce the detected attribute. They also observe a “