HATCorpus

Human–AI Text Corpus for Authorship Classification

Overview

HATCorpus is a curated dataset of human-written and AI-generated English text designed for research on authorship attribution, AI text detection, and stylometric analysis.

The dataset emphasizes clean labeling, transparent sourcing, and the absence of prompt leakage, making it suitable for benchmarking and academic research.

Primary use cases:

• Human vs AI text classification
• AI-generated content detection
• Stylometric and linguistic analysis
• Robustness and bias evaluation of detectors

---

Dataset Statistics

• Total samples: 32,248
• Language: English
• Labels:
  • 0 → Human-written
  • 1 → AI-generated

Label Distribution

Label	Count	Percentage
Human (0)	19,997	~62%
AI (1)	12,251	~38%

---

Data Sources

Human-Written Text

Human-authored samples are sourced from:

• Wikipedia (Jannuary 2022 dump)
• Project Gutenberg

These sources were selected to ensure authentic human authorship and permissive or public-domain licensing.

AI-Generated Text

AI-generated samples were produced using large language models with:

• Wikipedia-style prompts
• Thematic prompts across multiple domains

All AI-generated samples include their generation prompt for transparency and reproducibility.

---

Dataset Structure

Each row in the dataset contains the following fields:

Column	Description
text	The full text sample
label	0 = human, 1 = AI
source	Source category of the text
prompt	Prompt used for AI generation (NaN for human text)

Data Quality & Validation

The dataset was validated to ensure:

• No duplicate text samples
• No missing labels
• No prompt leakage into human-written samples
• Perfect consistency between label and source
• No explicit AI self-references (e.g., “As an AI model…”)

Known Characteristics & Limitations

Users should be aware of the following properties:

• AI-generated texts are longer on average than human-written texts
• Human text sources are primarily formal written corpora
• AI texts reflect high-quality, structured generations

Researchers are encouraged to control for text length and evaluate generalization on out-of-distribution data.

Intended Use

This dataset is intended for research and educational purposes, including:

• Training and evaluating AI text detection models
• Studying linguistic differences between human and AI writing
• Benchmarking authorship classification approaches

Ethical Considerations

• All human text sources are public-domain or permissively licensed
• AI-generated text does not represent real individuals
• The dataset does not claim AI-generated text is deceptive or inferior
• Detection models trained on this dataset may not generalize to all AI systems

Citation

If you use this dataset in your work, please cite:

@dataset{hatcorpus2025,
  title     = {HATCorpus: A Curated Human–AI Text Corpus for Authorship Classification},
  year      = {2025},
  publisher = {Hugging Face}
}