从直觉到专业：基于量规的认知校准提升人工识别LLM生成韩语文本的能力

Distinguishing human-written Korean text from fluent large language model (LLM) outputs remains challenging, even for linguistically trained readers who often over-rely on surface well-formedness. This research investigates whether expert-level detection can be treated as a learnable skill and enhanced through structured cognitive calibration. We introduce LREAD, a rubric derived from national Korean writing standards and specifically adapted to identify micro-level artifacts characteristic of LLM generation. These artifacts include, but are not limited to, optionality in punctuation, idiosyncratic spacing behaviors, selective particle application, unnatural sentence structures, and subtle semantic incoherence. Through a series of experimental trials, participants were guided to systematically evaluate Korean texts using the LREAD rubric. Experimental results demonstrate a significant improvement in the accuracy of identifying LLM-generated texts among evaluators who underwent rubric-based calibration. Furthermore, these calibrated evaluators were able to articulate clearer justifications for their judgments, moving beyond mere intuitive assessments. The LREAD rubric provides a standardized framework that redirects evaluators' attention from overall text fluency to deeper linguistic features and specific patterns indicative of LLM generation. This approach not only boosts detection accuracy but also enhances the confidence and explainability of evaluators' judgments. The study further analyzes the performance differences among evaluators of varying linguistic backgrounds and expertise levels after applying the LREAD rubric, revealing that systematic calibration training effectively improves discernment capabilities regardless of initial proficiency. This work offers crucial theoretical and practical foundations for developing more effective LLM-generated text detection tools and training methodologies in the future.