LLM Detection as an Intervention: Downstream Impact under Strategic User Behavior· 大模型检测作为干预：在战略性用户行为下的下游影响
As LLM adoption becomes more widespread, there is a growing interest in detecting LLM-generated content, for example through LLM detection tools and through heuristics based on language patterns. Detectors operate as an intervention that steers not only the detected attribute itself, but also downstream metrics such as LLM usage and output quality. In this work, we demonstrate how imperfect LLM detectors lead to counterintuitive impacts on these downstream metrics, by distorting how users are incentivized to use LLMs in their workflow. We develop a stylized model which captures how users strategically choose how much to use the LLM and how to post-process content to reduce the detected attribute. Using this model, we show that LLM detection can counterintuitively lead humans to increase th
研究大模型检测工具在用户策略性行为下的下游影响。
- 核心方法
- 建立了一个模型，模拟用户在使用大模型时如何战略性地选择使用程度和后处理内容以减少被检测的概率。
- 适合谁读
- 研究者、工程师、产品经理
- 要解决的问题
- 大模型检测工具的出现如何影响用户的使用行为及下游指标，如大模型的使用率和输出质量。
- 关键实验
- 通过模拟模型和案例分析展示了用户行为的变化及对下游指标的影响。
- 主要贡献
- 展示了大模型检测工具的不完美可能导致用户行为的反直觉变化，影响大模型的使用和输出质量。
- 意义与局限
- 揭示了大模型检测工具的实际应用中可能遇到的问题，为设计更有效的检测机制提供了理论支持。局限在于模型假设较为简化，实际应用中用户行为可能更为复杂。