RACE: Fine-grained AI-generated Text Detection via Dual-role Modeling of Creators and Editors

This article introduces the RACE (Rhetorical Analysis Modeling of Creators and Editors) method, which achieves fine-grained detection of four types of text—pure human text, pure AI text, AI-polished human text, and human-modified AI text—by distinguishing between the dual roles of creators and editors, providing a more precise technical solution for AI content regulation.

With the rapid development of large language models, the problem of AI-generated text abuse is becoming increasingly serious. However, most existing detection methods are limited to binary or ternary classification and cannot distinguish between the two scenarios of "AI-polished human-written text" and "human-modified AI-generated text". These two cases have completely different consequences in policy regulation, so a fine-grained detection solution is urgently needed.

The RACE method deconstructs text creation into two roles: "creator" and "editor". Different creation modes (pure human, pure AI, AI-polished human, human-modified AI) leave unique "signatures" on these two roles, enabling precise distinction through separate feature modeling.

RACE adopts two technical approaches: 1. Creator role: Introduce Rhetorical Structure Theory (RST) to build text logic diagrams, capturing systematic differences between humans and AI in argument organization; 2. Editor role: Focus on feature extraction at the level of Elementary Discourse Units (EDU), capturing stylistic traces such as the editor's vocabulary choices and sentence structure adjustments.

RACE was compared with 12 mainstream methods on multiple benchmark datasets, and it performed excellently in the four-classification task, especially in distinguishing the two complex scenarios of "AI-polished human text" and "human-modified AI text". It also maintains a low false positive rate, making it suitable for practical deployment.

The fine-grained detection capability of RACE can help implement regulations such as the EU AI Act and China's Interim Measures for the Management of Generative AI Services, assisting platforms in implementing differentiated management strategies (e.g., lenient management for AI-polished human content and strict labeling for human-modified AI content).

RACE relies on the accuracy of rhetorical structure parsing and EDU segmentation, and needs to address new generation modes from AI model iterations. In the future, it will be extended to multilingual and multimodal scenarios, and more robust feature extraction methods will be developed to deal with adversarial attacks.

RACE breaks through the limitations of traditional classification through dual-role modeling, providing a feasible technical path for fine-grained AI content regulation, and will become an important infrastructure for maintaining the health of the information ecosystem.