intermediate

8 min read

•Tuesday, April 7, 2026

Beyond the Bot: Unmasking the Creator and Editor in Hybrid AI Text

In an era where AI and humans co-author content, simply knowing if text is 'AI-generated' isn't enough. This groundbreaking research introduces a four-class detection system that discerns the subtle dance between human and AI, revealing who truly created the core idea and who merely polished it. For developers building with LLMs, this is the key to unlocking true content authenticity and policy-aligned AI applications.

Original paper: 2604.04932v1

Authors:Yang LiQiang ShengZhengjia WangYehan YangDanding Wang+1 more

Key Takeaways

1. Fine-grained LLM text detection moves beyond binary classification to a four-class system, distinguishing between human vs. LLM creators and editors.
2. RACE (Rhetorical Analysis for Creator-Editor Modeling) uses Rhetorical Structure Theory (RST) to identify the creator's high-level logical foundation and Elementary Discourse Unit (EDU) features for the editor's granular style.
3. This research is critical for establishing trust, ensuring compliance with emerging AI policies, and enabling sophisticated content moderation.
4. The ability to identify creator vs. editor roles provides invaluable insights for orchestrating multi-agent LLM systems and improving model fine-tuning.
5. The methodology offers practical applications for building advanced tools in academic integrity, content marketing, and legal document verification.

# The Paper in 60 Seconds

Imagine a world where you don't just know if a piece of text was written by an AI or a human, but you know *who had the core idea* and *who did the final edits*. That's precisely what the paper "Beyond the Final Actor: Modeling the Dual Roles of Creator and Editor for Fine-Grained LLM-Generated Text Detection" achieves.

Traditional LLM detection tools offer a binary (human/AI) or ternary (human/AI/collaborative) view. But the real world is more complex. This research introduces a four-class system:

1.Pure Human: Entirely written by a human.

2.Pure LLM: Entirely generated by an LLM.

3.LLM-polished Human: A human wrote the core content, and an LLM refined it (Human Creator, LLM Editor).

4.Humanized LLM: An LLM generated the core content, and a human refined it (LLM Creator, Human Editor).

The authors propose RACE (Rhetorical Analysis for Creator-Editor Modeling), a method that disentangles the 'creator's foundation' from the 'editor's style'. It uses Rhetorical Structure Theory (RST) to map the creator's logical flow and extracts Elementary Discourse Unit (EDU)-level features to capture the editor's granular stylistic choices. The result? Unprecedented precision in identifying the true origin story of text, with low false alarms.

Why This Matters for Developers and AI Builders

As AI agents become ubiquitous, the line between human and machine output blurs. For us building the next generation of AI-powered applications, understanding this nuance isn't just academic; it's critical for trust, compliance, and effective agent orchestration.

Think about it: Your AI agent might generate a draft, which a human then refines. Or a human might write an initial concept, and an LLM polishes it. The policy implications, ethical considerations, and even the user experience for these two scenarios are vastly different.

• Authenticity & Trust: In an age of deepfakes and AI-generated content, users demand transparency. Knowing the true authorship can build or break trust in platforms, news sources, or academic institutions.

• Policy & Compliance: Regulatory bodies are rapidly developing guidelines for AI-generated content. Distinguishing between 'LLM-polished human' and 'humanized LLM' is crucial for adhering to policies on academic integrity, journalistic standards, content moderation, or even legal document generation. A human-created, AI-polished research paper might be acceptable; an AI-created, human-polished one might not.

• Advanced AI Agent Orchestration: For companies like Soshilabs, orchestrating complex multi-agent systems is our bread and butter. If one agent acts as a 'creator' (e.g., a planning agent generating a core strategy) and another as an 'editor' (e.g., a refinement agent articulating that strategy), this research offers a powerful analytical lens. It allows us to understand *which agent contributed what* at a fundamental level, enabling better debugging, performance optimization, and attribution within agent workflows.

• Fine-tuning & Evaluation: By understanding how humans modify LLM output (and vice-versa), developers can gain invaluable insights into improving their models. What stylistic elements do humans consistently change? What logical flaws do LLMs introduce that require human correction? This data can inform more targeted fine-tuning and more robust evaluation metrics.

This isn't just about detection; it's about disentangling the collaborative process itself, offering a blueprint for more responsible and intelligent AI integration.

What the Paper Found: The Creator's Logic vs. The Editor's Style

The core innovation of this paper lies in its ability to model the dual roles of creator and editor. It moves beyond superficial textual cues to analyze deeper structural and stylistic signatures.

The Four-Class Spectrum

The authors rigorously define the four classes, acknowledging that 'hybrid' text isn't a monolith. This distinction is vital:

• LLM-polished Human Text: Here, the *idea* and *macro-structure* originate from a human. The LLM's role is refinement – improving clarity, grammar, flow, or rephrasing for impact. Think of a writer using Grammarly on steroids, or a PR professional getting an LLM to tighten up a press release they drafted.

• Humanized LLM Text: In this scenario, the *initial concept* and *bulk of the content* come from an LLM. A human then steps in to inject personality, correct factual errors, or adapt the tone to a specific audience. This is common when using LLMs for first drafts of marketing copy or blog posts, which then undergo human review and personalization.

RACE: Rhetorical Analysis for Creator-Editor Modeling

To differentiate these nuanced classes, RACE employs two distinct analytical lenses:

1.Rhetorical Structure Theory (RST) for the Creator's Foundation: RST is a linguistic theory that analyzes the hierarchical, functional relationships between parts of a text. It helps map out the *logic*, *argumentation*, and *coherence* of a document. For example, it identifies relationships like 'evidence-claim', 'problem-solution', or 'background-purpose'. The paper posits that the creator's signature is most evident in this high-level, structural organization. A human creator will likely establish a different logical flow or argumentative structure than an LLM creator, even if the subsequent editing is done by the other entity.

2.Elementary Discourse Unit (EDU)-level Features for the Editor's Style: EDUs are the smallest units of discourse that can be assigned a rhetorical relation (e.g., a clause or a short sentence). Analyzing features at this granular level—such as specific word choices, sentence complexity, syntactic patterns, or local coherence—reveals the editor's stylistic imprint. An LLM editor might introduce certain predictable phrasings or grammatical structures, while a human editor might inject more varied vocabulary or idiomatic expressions.

By combining these two perspectives, RACE can effectively disentangle who laid the conceptual groundwork (creator) and who shaped the linguistic surface (editor). The experiments demonstrate that this dual-modeling approach significantly outperforms existing baselines, achieving high accuracy with a critical low rate of false alarms, making it robust for real-world policy application.

How to Apply This: Building with Fine-Grained Detection

This research isn't just theoretical; it offers concrete mechanisms for building more intelligent and accountable AI systems. Here's what you could build:

• Advanced Content Moderation Systems: Move beyond simple spam detection. A platform could automatically flag 'pure LLM' content for immediate review, route 'humanized LLM' content to human moderators for factual verification, and allow 'LLM-polished human' content to pass with a 'AI-assisted' label. This creates a tiered, policy-aligned moderation pipeline.

• Intelligent Academic Integrity Checkers: Universities could deploy tools that don't just detect 'AI plagiarism' but differentiate between a student using an LLM for brainstorming and polishing (acceptable in some contexts) versus an LLM generating the entire essay with minor human tweaks (unacceptable). This fosters nuanced academic honesty.

• AI Agent Workflow Analytics for Soshilabs: Imagine a dashboard for your multi-agent system. When an agent chain produces a report, this technology could visually represent the 'creator-editor' split. Was the core strategy developed by the 'planning agent' (human creator) and merely articulated by the 'language generation agent' (LLM editor)? Or did the 'research agent' (LLM creator) generate the bulk of the content, which was then refined by a 'human review agent'? This provides invaluable debugging and performance insights.

• Dynamic Content Personalization: By analyzing how users edit LLM-generated content (i.e., acting as human editors), platforms can learn their specific stylistic preferences. This feedback loop could then be used to fine-tune LLMs to generate content that's 'closer' to the user's desired final output, reducing subsequent editing effort.

• Legal & Compliance Document Verification: In fields where precision and accountability are paramount, this could verify the true origin of legal briefs, contracts, or regulatory submissions. Ensuring a human lawyer conceptualized the core arguments, even if an LLM assisted in drafting, adds a critical layer of trust and legal defensibility.

This research offers a powerful lens through which to view human-AI collaboration, enabling us to build systems that are not only more intelligent but also more transparent, trustworthy, and aligned with societal expectations.

Key Takeaways

• Traditional LLM text detection is insufficient for the nuanced realities of human-AI collaboration.

• The paper introduces a four-class detection system (Pure Human, Pure LLM, LLM-polished Human, Humanized LLM) to differentiate creator and editor roles.

• RACE disentangles the 'creator's foundation' (using Rhetorical Structure Theory) from the 'editor's style' (using Elementary Discourse Unit-level features).

• This fine-grained analysis is crucial for policy-aligned AI regulation, content authenticity, and advanced AI agent orchestration.

• The technology enables building more sophisticated tools for content moderation, academic integrity, and workflow analysis in multi-agent systems.

Cross-Industry Applications

DevTools & AI Agent Orchestration

Implementing 'authorship attribution' within multi-agent LLM workflows to track which agent (or human) contributed the core idea vs. the final polish.

Significantly improves debugging, performance optimization, and accountability in complex AI agent systems, especially for Soshilabs' core business.

Education & Academic Integrity

Developing next-generation plagiarism detection tools that differentiate between a student using an LLM for concept refinement (Human Creator, LLM Editor) versus generating the entire essay with minor human edits (LLM Creator, Human Editor).

Fosters a more nuanced approach to academic honesty, allowing for appropriate use of AI tools while upholding integrity standards.

Content Marketing & Journalism

Automated analysis of articles and marketing copy to verify if content truly originates from a human ideator with AI assistance, or if it's primarily AI-generated with superficial human 'humanization'.

Enhances brand authenticity, builds reader trust, and informs content strategy based on the true human-AI collaboration ratio.

Legal & Compliance

Tools for verifying the origin and modification layers of legal documents, contracts, and regulatory submissions, ensuring that core legal arguments and concepts were human-conceived even if LLMs assisted in drafting.

Strengthens accountability and trust in legal documentation, reducing risks associated with purely AI-generated critical texts.

Back to Research Lab Read full paper