Beyond Binary Detection: Why Provenance Matters More Than Origin in LLM Text Classification

The detection of synthetic text has reached an inflection point where technical capability must confront regulatory nuance. While the field has spent years refining binary classifiers to distinguish human writing from machine generation, the reality of human-LLM collaboration has rendered such coarse distinctions increasingly untenable. The paper "Beyond the Final Actor: Modeling the Dual Roles of Creator and Editor for Fine-Grained LLM-Generated Text Detection" by Li et al. presents a compelling argument for abandoning binary and even ternary classification schemes in favor of a rigorous four-class taxonomy that separates LLM-polished human text from humanized LLM text. This distinction is not merely academic; it carries significant implications for academic integrity policies, content moderation frameworks, and our understanding of textual provenance.

The Limitations of Binary and Ternary Frameworks

Current detection systems predominantly operate under binary assumptions, categorizing text as either entirely human or entirely machine-generated. Some recent work has expanded to ternary classification, introducing a "collaborative" or "mixed" category for texts that show signs of human-LLM interaction. However, as Li et al. demonstrate, this aggregation conflates two fundamentally different creative processes with divergent ethical and policy consequences.

Consider the scenario of a student submitting an essay. If the student writes a draft and uses an LLM to polish grammar and improve readability, the resulting text originates from human conceptualization and logical organization, refined by machine stylistic optimization. Conversely, if the student prompts an LLM to generate the entire essay and subsequently humanizes the output to evade detection, the foundational structure emerges from machine reasoning, with human intervention limited to surface-level perturbations. Under most academic integrity policies, the former constitutes legitimate writing assistance while the latter represents academic dishonesty. Yet binary and ternary classifiers would likely assign both texts to the same category, or worse, misclassify the polished human text as purely machine-generated.

Li et al. formalize this distinction through a Creator-Editor framework. The creator establishes the foundational elements, logical flow, and argumentative structure, while the editor controls linguistic expression and surface-level style. In pure human or pure LLM text, these roles collapse into a single agent. However, in collaborative texts, the sequencing of creator and editor roles produces distinct forensic signatures that unified feature extraction cannot capture.

Technical Architecture: Rhetorical Structure and Discourse Units

To operationalize the Creator-Editor distinction, the authors propose RACE (Rhetorical Analysis for Creator-Editor Modeling), which explicitly separates the modeling of creation logic from editing style. The technical approach relies on two complementary linguistic theories.

For modeling the creator's contribution, RACE utilizes Rhetorical Structure Theory (RST), which parses text into hierarchical structures of rhetorical relations between Elementary Discourse Units (EDUs). These relations, such as elaboration, contrast, and cause, reflect the deep logical organization of the text. The authors construct an EDU-based logical relation graph where nodes represent discourse units and edges represent rhetorical dependencies. Through rhetoric-guided message passing, the model propagates information across this graph to capture complex argumentative dependencies, ultimately producing a root pooling representation that encodes the creator's structural fingerprint.

For the editor's contribution, RACE extracts EDU-level semantic representations that capture surface-level linguistic choices. These features reflect stylistic refinements, lexical substitutions, and syntactic variations that occur during the editing process without altering the underlying rhetorical structure.

The empirical justification for this decomposition appears in Table 1 of the paper, which shows sample-level cosine similarities between rhetorical-relation frequency vectors. Human-written text and LLM-polished human text exhibit high similarity (0.92), indicating that polishing preserves the original rhetorical structure created by the human. Conversely, LLM-generated text and humanized LLM text show even higher similarity (0.95), suggesting that humanization primarily perturbs surface features while leaving the machine's underlying logical organization intact. These patterns confirm that creator identity persists in rhetorical structure while editor identity manifests in discourse-level stylistics.

Policy-Aligned Detection and the Cost of Errors

Beyond the technical architecture, the paper addresses the practical requirements of regulatory deployment. In academic and professional contexts, false alarms carry significant costs. Misclassifying legitimate human writing as machine-generated can damage reputations and trigger unnecessary investigations, while failing to detect evasion undermines integrity systems.

RACE addresses these constraints by maintaining low false alarm rates while improving detection of the nuanced mixed classes. By looking beyond the final actor and reconstructing the creation pipeline, the method reduces the ambiguity that causes binary classifiers to err on collaborative texts. This precision matters for policy alignment; a detection system that cannot distinguish assistance from evasion cannot support fair enforcement.

The experimental results demonstrate RACE's superiority over 12 baseline methods under the four-class setting. The performance gains are particularly pronounced in separating the two mixed classes, where conventional approaches that rely on unified stylistic features or perplexity-based metrics fail to capture the directional asymmetry of creator-editor collaboration.

Original Insights: Toward Provenance Modeling

The contribution of Li et al. extends beyond improved classification accuracy. Their work signals a fundamental shift in how we conceptualize synthetic text detection. The field has historically treated detection as origin classification, asking simply "Was this written by a human or a machine?" The four-class framework and Creator-Editor decomposition suggest we should instead ask "Who created the foundation, who edited the expression, and in what sequence?"

This shift from origin to provenance has profound implications. First, it suggests that watermarking schemes, which embed signals at generation time, may prove insufficient for detecting humanized text where the original machine generation has been obscured. Forensic methods that reconstruct the creative process through structural analysis may become necessary complements to watermarking.

Second, the reliance on RST parsing introduces computational constraints that may limit real-time application. Rhetorical structure parsing remains computationally intensive and may struggle with informal genres or languages with limited discourse parsers. Future work must address whether lightweight approximations of rhetorical structure can maintain detection accuracy while improving throughput.

Third, the framework raises questions about attribution in iterative editing scenarios. If a text passes through multiple rounds of human and machine editing, does the creator-editor dichotomy remain binary, or do we require more complex graph structures representing multi-agent revision histories? The current four-class setting assumes a single creator and single editor, but real-world collaborative writing often involves more complex chains of custody.

Conclusion and Open Questions

The work of Li et al. establishes that effective LLM text detection requires moving beyond the final actor to model the full creative pipeline. Their RACE method demonstrates that rhetorical structure analysis captures creator signatures that persist through editing, while discourse-unit features isolate editor contributions. This provenance-aware approach aligns technical detection capabilities with regulatory requirements, distinguishing legitimate assistance from deceptive evasion.

Looking forward, several questions remain open. How does this framework extend to multimodal content where creation and editing involve images, audio, or video? Can we develop similar structural analyses for code, where creator logic manifests in algorithmic architecture and editing appears in variable naming and formatting? And as LLMs become capable of mimicking not just style but argumentative structure, will rhetorical features remain discriminative, or must we develop deeper semantic fingerprints?

The path forward likely involves hybrid systems that combine watermarking, structural analysis, and behavioral profiling. The binary classification era served its purpose for early detection challenges, but the future belongs to fine-grained provenance modeling that respects the complexity of human-machine collaboration. The question is no longer whether machines can write like humans, but whether we can reconstruct the history of how a text came to be.