Beyond Binary: Unmasking the Creator and Editor in LLM-Generated Text

The Paper in 60 Seconds

The Problem: Current LLM detectors are largely binary (pure human vs. pure AI) or, at best, ternary (adding a 'collaborative' category). This isn't enough for the real world. Policy decisions often hinge on *how* AI was used. Is it a human's idea polished by an LLM, or an LLM's idea humanized for deception?

The Solution: The paper introduces RACE (Rhetorical Analysis for Creator-Editor Modeling), a novel method for fine-grained LLM detection across a rigorous four-class setting:

How it Works: RACE is groundbreaking because it separates the *creator's* signature from the *editor's* signature. It uses Rhetorical Structure Theory (RST) to build a logic graph that reveals the creator's foundational structure, and extracts Elementary Discourse Unit (EDU)-level features to identify the editor's stylistic choices.

The Impact: RACE outperforms existing baselines, offering a policy-aligned solution for LLM regulation with significantly lower false alarms. This means developers can build tools that understand the true provenance of text.

Why This Matters for Developers and AI Builders

The era of simple 'AI-or-not' detection is rapidly coming to an end. As Large Language Models (LLMs) become seamlessly integrated into every facet of content creation – from drafting emails and generating code to summarizing reports and crafting marketing copy – the line between human and machine output is not just blurring; it's becoming a complex, multi-layered tapestry. For developers and AI builders, this complexity isn't just an academic curiosity; it's a critical challenge demanding sophisticated solutions.

Consider the implications:

Both scenarios involve human and LLM collaboration, but their provenance and intent are vastly different, triggering distinct policy consequences in areas like academic integrity, journalistic ethics, legal compliance, and social media moderation. Building responsible AI means understanding not just *if* text is AI-generated, but *how* it was generated and by whom. This research provides the framework for building the next generation of tools that can make these crucial distinctions, fostering trust, preventing misuse, and enabling targeted interventions.

Beyond the Binary: The Four-Class Revolution

Traditional LLM detection methods often fall short because they operate on an oversimplified understanding of human-AI collaboration. A binary classifier can only tell you 'human' or 'AI.' A ternary system might add a 'mixed' or 'collaborative' category, but even this lumps together fundamentally different scenarios. The key insight of this paper is that LLM-polished human text (HL) and humanized LLM text (LH) are not the same, and treating them as such leads to flawed policies and ineffective moderation.

RACE's four-class setting provides the necessary granularity:

This nuanced classification empowers developers to build systems that reflect the true complexity of modern content creation and respond appropriately to each scenario.

RACE: Unpacking Creator and Editor Signatures

The brilliance of RACE lies in its ability to disentangle the distinct 'signatures' of the creator and the editor, even when both are involved and one of them is an LLM. The core hypothesis is that LLMs and humans leave different traces when they *originate* ideas (the creator role) versus when they *refine* existing text (the editor role).

Creator's Signature via Rhetorical Structure Theory (RST)

To identify the creator's foundation, RACE leverages Rhetorical Structure Theory (RST). RST is a linguistic framework that analyzes text organization by identifying rhetorical relations (e.g., cause-effect, elaboration, contrast, background) between segments of text. Imagine a text as a tree structure, where leaves are elementary units and branches represent how these units logically connect to form larger ideas.

Editor's Signature via Elementary Discourse Units (EDUs)

To identify the editor's style, RACE focuses on Elementary Discourse Units (EDUs). EDUs are the smallest meaningful units of discourse, roughly corresponding to clauses. The editor's role is often about refining language, improving flow, and adjusting tone – changes that manifest at a finer textual granularity.

By combining these two distinct analytical layers, RACE can accurately untangle who originated the core message and who shaped its final presentation, providing unparalleled insights into text provenance.

What Can You BUILD with RACE? Practical Applications for Developers

The power of RACE extends far beyond academic research. For developers, this fine-grained detection capability opens up a new frontier for building more intelligent, ethical, and robust AI-powered applications.

The Future is Fine-Grained

As LLMs continue to evolve and become indistinguishable from human writers, the need for sophisticated, fine-grained detection will only grow. RACE offers a powerful, policy-aligned solution that moves us beyond simplistic binary judgments. By understanding the distinct roles of the creator and editor, whether human or AI, developers are now equipped to build the next generation of transparent, trustworthy, and ethically sound AI systems. The future of AI content is not just about generation; it's about intelligent, nuanced understanding of its origins.