AI-edited text was more often mistaken for human writing than for fully AI-generated text in a new preprint study, suggesting that editing and full generation leave different detection problems. In a program that sorted passages into human-written, AI-generated and AI-edited categories, edited passages were labeled human 31.9% of the time, while only 5.5% were labeled AI-generated. The same three-way test reached 75.4% overall accuracy. In the study’s measured style features, edited text therefore sat closer to human writing than to text generated from scratch.
What the researchers tested
The findings come from an arXiv version 1 preprint dated 28 Aug. 2026. The study asks whether AI-generated writing has a consistent stylometric style, meaning a measurable pattern in how text is written, across models and domains, and whether AI-edited writing follows the same pattern.
It describes 45,000 human-written and AI-generated texts across eight language models and five domains, plus 273,420 AI-edited texts produced with three models, 31 prompts and eight editing categories.
For the editing arm, the researchers started with 200 human-written documents per domain, 1,000 in total, and set target edit ratios of 30%, 50% and 70%. They assessed each output against its source with a word-level Levenshtein edit ratio, a measure of how much the wording differed.
The basic signal was real, but uneven
To make style measurable, the feature pipeline began with 26 candidate measures and retained 14 after removing measures that overlapped too strongly. The retained features were calculated from each passage’s raw text, without access to a language model, external API calls or corpus-level statistics. The main classifier was logistic regression, a standard method for sorting examples into categories. It used balanced sampling, stratified 80%/20% training and test splits, and ten independent runs. Feature importance was assessed from model coefficients and from the drop in held-out accuracy when a feature was shuffled.
On the human-versus-AI-generated task, the stylometric classifier reached a mean accuracy of 91.74%, with a run-to-run standard deviation of 0.61 percentage points. Accuracy ranged from 92.57% to 95.83% across domains, but from 79.45% to 99.09% across the listed language models. The broad separation was therefore consistent across the tested material, but not equally strong for every generator.
Two measures formed the most stable part of the signal. AI-generated documents had mean entropy, a measure of variation in text patterns, about one standard deviation higher than human documents, and lexical diversity, the range of different words used, about two standard deviations higher. Lexical diversity ranked among the five most important features in 13 of 14 conditions, and entropy did so in 12 of 14. Other feature rankings changed by condition, and burstiness was not consistently important.
Editing did not simply repeat generation
Editing did not simply reproduce the full-generation pattern. Relative to the human source, AI editing increased lexical diversity by Cohen’s d of +0.24, decreased entropy by d of minus 0.28 and produced the largest overall shift in lexical density, at d of minus 3.10. Cohen’s d is a standardized way of expressing the gap between two groups, so these figures describe the size and direction of an association, not a cause.
Yet the average concealed differences among editors. With GPT-OSS 120B, entropy instead rose by d of +0.44, word count stayed roughly constant at d of +0.02, and the Linsear Write measure fell by d of minus 2.81. The reported effects also varied by editing model, category and domain.
Some checks found the separation pattern persisted when the generation temperature changed. Accuracy was 82.2% at temperature 0.3, 80.7% at 0.7 and 79.5% at 1.0. In a targeted proprietary-generator subset, accuracy was 82.1% for Gemini-3-Flash and 78.6% for GPT-5.4-mini. The proprietary analysis was smaller than the main benchmark, so it is supporting evidence rather than a like-for-like replacement for the main result.
One detector was not enough
The contrast was clearest when the study compared the two kinds of AI text directly. Stylometric features produced an AUC of 0.98 for AI-edited versus AI-generated text. AUC is the study’s measure of how well a score separates two groups. The same features reached 0.80 for human versus AI-edited text, while the editing-aware system EditLens reached 0.97 on that comparison.
Combining the stylometric classifier with EditLens-L raised the two pairwise results to 0.99 for AI-edited versus AI-generated text and 0.94 for human versus AI-edited text. In this corpus, using both kinds of signal gave strong separation in both tasks.
When all three categories were tested together, the classifier got 75.4% right and correctly identified 89.2% of AI-generated texts. AI-edited texts were misclassified as human 31.9% of the time, compared with 5.5% as AI-generated. The confusion pattern places edited text closer to human than to fully generated writing in this test.
What the test could not settle
The result has clear boundaries. The feature set was intentionally limited to lightweight, interpretable stylometric measures. The generators and domains were not exhaustive, large-scale editing used only a small set of open-weight editors, and the documents were mainly English long-form text. The study also warns that these feature associations do not show that any one feature caused the separation.
Within those boundaries, the preprint’s central message is narrower than a claim about all AI writing: a detector that separates human from fully generated text may see edited text differently, while combining stylometric with editing-aware signals may help distinguish the categories.
Paper data and sources
Original title: AI Writers Have a Consistent Stylometric Footprint, but AI Editors Do Not
Authors: Zhengyang Shan, Yukyung Lee, Sophie Hao
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text