Entropy and lexical diversity flag AI writing at 91.7%; AI editing cuts density instead

AI Writers Have a Consistent Stylometric Footprint, but AI Editors Do Not

Zhengyang Shan, Yukyung Lee, Sophie Hao

EMNLP Main 2026

cs.CL

2026-08-28

AI writing has a stable entropy-and-diversity footprint (91.74% linear acc. across 8 models, 5 domains). AI editing cuts lexical density (d=-3.10) and stays near the human source.

What problem this solves

AI-text detectors assume a shared statistical stain: if a model touched the document, the same features should fire, whether the model wrote from scratch or edited a human draft. Detectors trained on fully generated text then get pointed at copy-edits, paraphrases, and expansions. That transfer was never tested on one compact, interpretable feature set.

Boston University asks two concrete questions. Does AI-generated text have a style that holds across models and domains? Do AI edits follow it?

Method

They build GEN (Generated, Edited, and Naturalistic) in five English domains: Wikipedia, WikiHow, arXiv abstracts, Reddit QA, and Reddit stories. Human documents: 4,998, mostly from M4 and WritingPrompts. AI-generated: 39,916, written by eight open instruction models (Qwen-2.5 7B/72B, Gemma-3 12B/27B, Llama-3.1-8B, Llama-3.3-70B, GPT-oss-20B/120B). AI-edited: 273,420, from 1,000 human seeds, 31 prompts, eight edit types (grammar, fluency, clarity, tone, paraphrase, add detail, restructure, condense), three large editors, target edit ratios 30%/50%/70%. Actual edit ratio is word-level Levenshtein, measured after the fact.

Each document gets 14 interpretable features: lexical diversity and density, entropy, burstiness, readability, punctuation, parse-tree depth, and related surface stats. The detector is L2-regularized logistic regression on standardized features, averaged over ten class-balanced draws. Importance is both the absolute coefficient and the accuracy drop after shuffling that feature.

Results

Generation has a thin, stable footprint. Versus human text, AI-generated documents sit about 1 standard deviation higher in entropy and about 2 standard deviations higher in lexical diversity. With only these 14 dimensions, human vs generated accuracy is 91.74%±0.61 globally, 92.57%–95.83% per domain, and 92%–99% for most generators. GPT-oss-120B is the outlier at 79.45%, the most human-like generator in the set.

Across 14 training conditions (global, per-domain, per-model), lexical diversity ranks in the top five 13 times and entropy 12 times; both have the highest stability scores. Burstiness ranks third globally and then fails to travel. Temperature 0.3/0.7/1.0 yields 82.2%/80.7%/79.5% accuracy, with mean Spearman ρ 0.89 on feature ranks. Proprietary add-on: Gemini-3-Flash 82.1%, GPT-5.4-mini 78.6%. Lexical diversity still helps; secondary features already change.

Editing takes a different path. Versus the human source, lexical diversity rises only slightly (Cohen's d=+0.24) and entropy falls (d=-0.28). The generation pattern of both going up does not appear. The dominant edit effect is a collapse in lexical density (d=-3.10): content-word share drops, function words rise. That same density feature had permutation importance of only 0.056 for generation detection. GPT-oss-120B as an editor is again the odd one out, raising entropy (d=+0.44) and cutting Linsear Write hard (d=-2.81).

Pairwise AUC:

MethodHuman vs genEdit vs genHuman vs edit
Stylometric LR0.970.980.80
EditLens-Llama1.000.850.97
Score average1.000.990.94

Stylometric features separate edited from generated text and struggle to separate edited from human text. EditLens does the reverse. At 5% false-positive rate, EditLens-Llama's true-positive rate for human vs edit is 91%. Edited text hugs the source, not the generation cloud. It is not a midpoint.

Why it matters

"AI text" is not one label. A detector trained on fully generated documents will systematically miss "human-written, model-edited" drafts if it relies on this stylometric footprint: those drafts hug the source, not the generation cloud. Classroom checkers, journal screens, and platform labels that train only on pure generation will under-flag polish, paraphrase, and expansion. Stylometric models and neural edit detectors see different signals; combining them covers both jobs. This is a measurement result, not a new leaderboard detector.

Limitations

Fourteen features were chosen to stay interpretable and cheap. Semantics, argument structure, and factual hallucination are outside the set. The main study uses eight open models; editing uses only three 70B-class open models; proprietary generation is a smaller add-on of two models. The corpus is English long-form. Chat, multilingual, and highly technical writing are untested. Feature importance is associational: high entropy need not be "because it is AI," it can travel with length, domain, and decoding. Edit prompts come from the EditLens suite; real user edits may look nothing like them. GPT-oss-120B already splits from the other models on both generation and editing, so the footprint is starting to crack by model family.

Terms

Source

What people are saying

Related papers

All paper explainers