Human, AI-Generated, and AI-Edited Text: Stylometric Corpus

📄 Paper: https://arxiv.org/pdf/2608.27855

💻 Code: https://github.com/ZhengyangShan/GEN-stylometric-footprint

A three-class corpus for studying how AI writing differs from human writing, distinguishing two modes of AI involvement:

• AI generation: text written by an LLM from scratch, given a prompt.
• AI editing: human text revised by an LLM (grammar, tone, paraphrase, etc.).

Supports detection of AI-generated text, AI-edited text, and the 3-class distinction between human / AI-generated / AI-edited writing. Every text carries a prompt_id so a human reference, its AI generations, and its AI edits can be joined into one group.

Classes

label	text_type	count	description
0	human	5,000	human-written references (1,000 / domain)
1	ai_generated	39,391	LLM-written from scratch (8 models)
2	ai_edited	273,418	human text edited by an LLM (3 models x 8 edit types x 3 ratios)

Files

File	Class	Rows
human.jsonl	human (0)	5,000
ai_generated.jsonl	ai_generated (1)	39,391
ai_edited.jsonl	ai_edited (2)	273,418

Schema

All rows share a common core plus a join key; each class adds its own metadata.

Common (every row):

Field	Description
prompt_id	shared join key (0-4999): identifies the source text a row derives from
text	the text (cleaned for AI rows; unmodified for human)
label	0 = human, 1 = ai_generated, 2 = ai_edited
text_type	string form of the label
source	writing domain
model	generating / editing model (human for human text)

human / ai_generated also have:

• prompt: the generation instruction

ai_edited also has:

• seed_text: the human original that was edited
• seed_id: editing-corpus doc id (0-979); groups edits within the editing corpus
• edit_category: one of 8 editing types
• edit_prompt: the specific edit instruction
• edit_ratio: measured word-level Levenshtein edit ratio (%)
• target_ratio: target edit ratio (30 / 50 / 70)
• edit_ratio_bin: binned edit ratio

Joining on prompt_id

A single prompt_id gathers one human reference, its (up to 8) AI generations, and all its AI edits, forming the same underlying text across all three classes.

import json
from collections import defaultdict

groups = defaultdict(lambda: {"human": None, "ai_generated": [], "ai_edited": []})
for fn in ["human.jsonl", "ai_generated.jsonl", "ai_edited.jsonl"]:
    for line in open(fn):
        o = json.loads(line)
        g = groups[o["prompt_id"]]
        if o["text_type"] == "human":
            g["human"] = o
        else:
            g[o["text_type"]].append(o)

# Everything derived from source text #0:
g = groups[0]
print(g["human"]["text"][:80])              # the human reference
print(len(g["ai_generated"]), "AI generations")
print(len(g["ai_edited"]),   "AI edits")

Tasks

from datasets import load_dataset, concatenate_datasets
human = load_dataset("szyszy/GEN", "human", split="train")
gen   = load_dataset("szyszy/GEN", "ai_generated", split="train")
edit  = load_dataset("szyszy/GEN", "ai_edited", split="train")

common = ["prompt_id", "text", "label", "text_type", "source", "model"]
three_class = concatenate_datasets([
    human.select_columns(common), gen.select_columns(common), edit.select_columns(common)
])
# Binary tasks are label filters:
#   Human vs AI-gen   -> {0, 1}
#   AI-edit vs AI-gen -> {1, 2}
#   Human vs AI-edit  -> {0, 2}

Generation models (8)

Model	AI generations
gemma3-12B	4,954
gemma3-27B	5,011
gpt-oss-120B	4,985
gpt-oss-20B	4,451
llama3.1-8B	4,995
llama3.3-70B	4,999
qwen2.5-72B	4,997
qwen2.5-7B	4,999

Editing models (3)

qwen2.5-72B, llama3.3-70B, gpt-oss-120B: the three largest generation models, applied to revise human seed texts.

Edit categories (8): grammar_and_mechanics, fluency_and_flow, clarity_and_precision, tone_and_style, paraphrasing, adding_detail, structure_and_organization, concision.

Edit ratios: each seed is edited at three target ratios (30 %, 50 %, 70 %); the actual per-text edit ratio is in edit_ratio.

Domains (5)

Domain	Rows (generation)
arxiv	7,833
reddit	7,872
story_generation	7,966
wikihow	7,796
wikipedia	7,924

arxiv (scientific abstracts), reddit (Q&A / ELI5), story_generation (creative writing), wikihow (procedural), wikipedia (encyclopedic).

Generation & editing configuration

AI-generated text: temperature 0.7, top-p 0.9, max 1024 new tokens, matched to each human reference's prompt. AI-edited text: the three largest models applied 31 edit prompts across 8 categories at 3 target ratios to human seeds.

Cleaning

All AI text (generated and edited) was cleaned to remove model artifacts that would make detection trivial and unrepresentative of natural writing: markdown formatting, chat-template leakage, multi-option responses ("Option 1: ..."), meta-preambles ("Here's a rephrased version:", "Here's the edited text:"), trailing meta-commentary, refusals, and reasoning/planning leakage. Human references are unmodified. Natural conversational openers appropriate to a domain (e.g. a Reddit answer beginning "Sure! You know how...") are preserved.

Notes

• gpt-oss-20B generations are shorter on average; rows under 150 chars were dropped.
• Length varies by domain and class (human ArXiv abstracts are short, WikiHow articles long, AI text more uniform). Control for length where relevant.
• ai_edited is much larger than the other classes (273K vs 39K vs 5K). Subsample per model / edit_category / edit_ratio_bin for balanced experiments.

License

Released under CC-BY-NC-4.0. Human reference texts retain the licenses of their source corpora.

Citation

@misc{shan2026aiwritersconsistentstylometric,
      title={AI Writers Have a Consistent Stylometric Footprint, but AI Editors Do Not}, 
      author={Zhengyang Shan and Yukyung Lee and Sophie Hao},
      year={2026},
      eprint={2608.27855},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2608.27855}, 
}