Early large language model (LLM) detectors operated on a simple premise: a document is either entirely human-written or entirely machine-generated. This binary paradigm works well in controlled laboratory settings, but it fails in the wild.
In real-world workflows, humans rarely just copy-paste raw LLM outputs. Instead, they use AI as a co-pilot—alternating between human drafting and machine generation, or using models to iteratively polish their prose. As noted by Zhang et al., this collaborative process creates "mixtext"—hybrid content that blends human and machine contributions. When forced into a traditional binary classifier, hybrid text often lands in an ambiguous statistical middle ground, causing detection performance to plummet to near random guessing.
To secure content authenticity and academic integrity, a new research direction has emerged. Rather than asking whether a text is AI-generated, researchers are developing methods to detect where and to what extent an AI was involved.
Why the Binary Assumption Fails
Traditional AI text detectors rely on global statistical signatures, such as uniform perplexity, token cohesiveness, or entropy across a long sequence. When human and machine text are interleaved, these global signatures are disrupted.
The blending of authorship masks the predictable patterns typically found in raw machine-generated text in two distinct ways:
- AI Polishing Over Human Drafts: Even minimal AI intervention can completely alter a text's statistical profile. Saha et al. demonstrated that minor stylistic touch-ups by an LLM can trigger disproportionately high confidence scores from binary detectors, leading to false accusations of full machine generation.
- Human Editing Over AI Drafts: Conversely, when a human edits an AI-generated draft, they inject stylistic variations that mask the machine's underlying statistical footprint. As Srba observed, a single round of AI paraphrasing or human rewriting can effectively erase an author's original signature, creating a conceptual and technical ambiguity around who the primary "author" actually is.
To address these challenges, the field categorizes hybrid text into two structures: heterogeneous mixed text (distinct blocks or sentences written by different entities) and homogeneous mixed text (where AI has polished or edited human prose at a fine-grained level).
Detecting Heterogeneous Boundaries
For text composed of distinct human and machine segments, the solution lies in shifting from document-level classification to fine-grained sentence-level or paragraph-level localization.
Instead of averaging features across the entire document, modern frameworks look for the exact transition points where authorship shifts. Ippolito et al. pioneered the study of these transition points through the Real or Fake Text (RoFT) framework, demonstrating that while generative models can maintain stylistic consistency over short sequences, boundaries still exhibit detectable shifts.
[Human Sentence] → [Human Sentence] → 边界 (Shift in Style/Probability) → [AI Sentence]
To automate this boundary detection, Wang et al. introduced SeqXGPT. This framework treats word-wise log probabilities from language models as "wave-like" signals. By analyzing these signals through a hybrid architecture, it can pinpoint specific machine-generated segments within a larger human context.
The intuition here is that the interface between human-authored and AI-generated text creates local stylistic incoherence. As shown by Xu et al., context-aware detection models significantly outperform independent sentence analysis because they explicitly capture these abrupt statistical fluctuations at the boundaries where human and AI text meet.
Measuring the Gradient of Collaboration
When text is homogeneously mixed—such as an essay written by a human but completely polished by ChatGPT—clear boundaries do not exist. In these scenarios, forcing a discrete classification fails. The field has therefore shifted toward continuous estimation and role recognition.
Instead of a binary label, Yang et al. introduced the concept of a "Polish Ratio" to quantify the exact degree of AI intervention in a document. This approach recognizes that the statistical signatures of AI intervention exist on a spectrum depending on how heavily the text was modified.
Similarly, Zhou proposed moving toward fine-grained role recognition, designing systems that identify the specific role the LLM played (e.g., whether it acted as a minor polisher or the primary content creator) and measuring its total involvement.
This gradient perspective is crucial for practical applications. For example, in an educational setting, a detector must distinguish between permitted assistive usage (like grammar correction) and prohibited content generation. To make this distinction reliable, Ren introduced a conformal prediction framework that uses statistical calibration to set adaptive thresholds, allowing educators to catch full AI generation while safely ignoring minor, permitted AI touch-ups.