Fine-Grained Detection of Human-AI Collaborative Writing

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2275 papers574 conceptsfrozen 2026-10-04

The traditional "cat-and-mouse" game of AI text detection has long relied on a binary assumption: a document is either entirely human-written or entirely machine-generated. However, as Large Language Models (LLMs) become integrated into professional workflows, this dichotomy is collapsing. We are entering an era of "mixtext" or "cyborg" authorship, where humans prompt, edit, and weave AI suggestions into their own prose.

Recent research indicates that standard binary classifiers are ill-equipped for this reality. Han Xu and colleagues have shown that these models often overfit to irrelevant features like text length and struggle to generalize across different domains. To address this, a new research direction is emerging that focuses on fine-grained detection—identifying not just if AI was used, but how and to what extent.

The Spectrum of Authorship

The core challenge lies in the "grey area" of AI involvement. Mervat Abassy and Zihao Cheng argue for a shift toward multi-class taxonomies. Instead of a simple "yes/no" verdict, these frameworks distinguish between four distinct categories:

  1. Purely Human: Original text with no machine intervention.
  2. Purely Machine: Text generated entirely from a prompt.
  3. Human-Polished Machine: AI-generated text edited by a human to bypass detectors.
  4. Machine-Polished Human: Human-authored drafts refined by an LLM for grammar, flow, or style.

This last category—machine-polished human text—is particularly elusive. Because the underlying logic and structure remain human-driven, these texts often evade statistical markers used to identify fully synthetic content. Lingyi Yang introduced the Polish Ratio to quantify this specific type of involvement, providing a metric for how much an LLM has "touched" a human draft.

Quantifying Involvement and Effort

Building intuition for this requires moving beyond linguistic style to information theory. Yueqi XIE proposed a framework to measure human contribution by calculating the mutual information between human prompts and AI outputs. This allows researchers to distinguish between high-effort tasks (where a human provides a detailed draft for polishing) and low-effort tasks (where a human provides a single-word prompt for generation).

In the context of academic integrity, Weixin Liang developed statistical methods to estimate the "AI-generated fraction" of large datasets. Their analysis of recent AI conference peer reviews found that up to 17% of sentences showed signs of substantial LLM modification, particularly as deadlines approached. This suggests that AI is often used as a productivity aid rather than a total replacement for human expertise, a nuance that binary detectors miss.

Locating the Boundary

Another facet of this research focuses on boundary detection: identifying the exact "seam" where a human stops writing and an AI takes over. Daphne Ippolito and her team found that while humans struggle to spot these transitions, the task is essential for identifying where misinformation or logical inconsistencies might begin in a collaborative document.

This transition point is where the "Ship of Theseus" paradox applies to prose. As explored by Nafis Irtiza Tripto, even a single round of AI paraphrasing can effectively erase a human author's stylistic signature. If every "plank" of a sentence is replaced by an LLM's preferred phrasing, the resulting text may align more closely with the AI's statistical profile than the original author's, even if the core ideas remain human.

New Mechanisms for Detection

To catch these subtle interventions, researchers are moving toward more granular, multi-level analysis. Zhen Tao proposed the Multi-level Fine-grained Detection (MFD) framework, which analyzes text at the sentence level across statistical, semantic, and deep linguistic tiers.

Other innovative approaches include:

  • Indirect Prompt Injection: Vidya Rao and others have experimented with embedding hidden instructions in PDFs (like "include a fake citation") that force an LLM to leave a "digital breadcrumb" if it is used to summarize or review the document.
  • Implicit Reward Models (IRM): Runheng Liu developed a zero-shot detector that measures how well text aligns with human preferences, identifying the "aligned" signature of instruction-tuned models.

Why It Works

Fine-grained detection works because it acknowledges that AI intervention leaves a unique statistical footprint, even when it doesn't generate the core content. While a human provides the "what" (the logic and facts), the LLM provides the "how" (the linguistic distribution). By training models on datasets like MIXSET or CHEAT—which specifically contain blended human-AI text—detectors can learn to recognize the specific signatures of collaborative writing as a distinct category rather than just noise in a binary system.

Papers in this version

Showing 60 of 275 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 148 papers
  2. Version 1 · 2026-10-04 · 275 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_hybrid_authorship_detection_26450750,
  title = {hybrid authorship detection},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hybrid-authorship-detection/rdr_26450750754e4965bc264a6c97714bd4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_fine_grained_detection_of_human_ai_colla_26450750_v1,
  title = {Fine-Grained Detection of Human-AI Collaborative Writing},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/fine-grained-detection-of-human-ai-collaborative-writing/rdr_26450750754e4965bc264a6c97714bd4/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}