Detection of Machine-Generated Arabic Text in the Era of Large Language Models

This repository contains the official implementation and datasets for the research paper "The Arabic AI Fingerprint: Stylometric Analysis and Detection of Large Language Models Text" (https://arxiv.org/abs/2505.23276) by Maged S. Al-Shaibani and Moataz Ahmed.

📋 Overview

This work represents the first comprehensive investigation and stylometric analysis of Arabic machine-generated text detection across multiple llms and generation methods, addressing the critical challenge of distinguishing between human-written and AI-generated Arabic content across multiple domains and generation strategies.

🎯 Contributions

• Multi-dimensional stylometric analysis of human vs. machine-generated Arabic text
• Multi-prompt generation framework across 4 LLMs (ALLaM, Jais, Llama 3.1, GPT-4)
• High-performance detection systems achieving up to 99.9% F1-score
• Cross-domain evaluation (academic abstracts + social media)
• Cross-model generalization studies

🏗️ Repository Structure

arabic_datasets/
    ├── arabic_filtered_papers.json
    └── social_media_dataset.json
generated_arabic_datasets/
    ├── allam/
        ├── arabic_abstracts_dataset/
            ├── by_polishing_abstracts_abstracts_generation_filtered.jsonl
            ├── by_polishing_abstracts_abstracts_generation.jsonl
            ├── from_title_abstracts_generation_filtered.jsonl
            ├── from_title_abstracts_generation.jsonl
            ├── from_title_and_content_abstracts_generation_filtered.jsonl
            └── from_title_and_content_abstracts_generation.jsonl
        ├── arabic_social_media_dataset/
            ├── by_polishing_posts_generation_filtered.jsonl
            └── by_polishing_posts_generation.jsonl
        └── arasum/
            └── generated_articles_from_polishing.jsonl
    ├── claude/
        ├── arabic_abstracts_dataset/
            ├── by_polishing_abstracts_abstracts_generation.jsonl
            ├── from_title_abstracts_generation.jsonl
            └── from_title_and_content_abstracts_generation.jsonl
        └── arasum/
            └── generated_articles_from_polishing.jsonl
    ├── jais-batched/
        ├── arabic_abstracts_dataset/
            ├── by_polishing_abstracts_abstracts_generation_filtered.jsonl
            ├── by_polishing_abstracts_abstracts_generation.jsonl
            ├── from_title_abstracts_generation_filtered.jsonl
            ├── from_title_abstracts_generation.jsonl
            ├── from_title_and_content_abstracts_generation_filtered.jsonl
            └── from_title_and_content_abstracts_generation.jsonl
        ├── arabic_social_media_dataset/
            ├── by_polishing_posts_generation_filtered.jsonl
            └── by_polishing_posts_generation.jsonl
        └── arasum/
            └── generated_articles_from_polishing.jsonl
    ├── llama-batched/
        ├── arabic_abstracts_dataset/
            ├── by_polishing_abstracts_abstracts_generation_filtered.jsonl
            ├── by_polishing_abstracts_abstracts_generation.jsonl
            ├── from_title_abstracts_generation_filtered.jsonl
            ├── from_title_abstracts_generation.jsonl
            ├── from_title_and_content_abstracts_generation_filtered.jsonl
            └── from_title_and_content_abstracts_generation.jsonl
        ├── arabic_social_media_dataset/
            ├── by_polishing_posts_generation_filtered.jsonl
            └── by_polishing_posts_generation.jsonl
        └── arasum/
            └── generated_articles_from_polishing.jsonl
    └── openai/
        ├── arabic_abstracts_dataset/
            ├── by_polishing_abstracts_abstracts_generation_filtered.jsonl
            ├── by_polishing_abstracts_abstracts_generation.jsonl
            ├── from_title_abstracts_generation_filtered.jsonl
            ├── from_title_abstracts_generation.jsonl
            ├── from_title_and_content_abstracts_generation_filtered.jsonl
            └── from_title_and_content_abstracts_generation.jsonl
        ├── arabic_social_media_dataset/
            ├── by_polishing_posts_generation_filtered.jsonl
            └── by_polishing_posts_generation.jsonl
        └── arasum/
            └── generated_articles_from_polishing.jsonl
hf_export/
    ├── abstracts_dataset.py
    └── social_media_dataset.py
models/
    ├── __init__.py
    ├── data.py
    ├── models.py
    └── train.py
notebooks/
    ├── Arabic_experiments/
        ├── ArabicAbstractsDataset/
            ├── Arabic_abstracts_dataset_preparation.ipynb
            ├── continual_training_of_arasum_detector_on_arabic_abstracts.ipynb
            ├── llms_multi_class_arabic_detector.ipynb
            ├── train_on_one_model_test_on_others.ipynb
            ├── train_on_one_prompt_test_on_others.ipynb
            └── zero_shot_on_arabic_abstracts_dataset.ipynb
        ├── ArabicSocialMediaDataset/
            ├── llms_multi_class_arabic_detector.ipynb
            ├── prepare_arabic_social_media_dataset.ipynb
            └── train_on_one_model_test_on_others.ipynb
        └── AraSum/
            ├── AllamWithAraSumTestingOnly.ipynb
            ├── arabic_detector_trained_on_all_llms.ipynb
            ├── arabic_detector_trained_on_allam.ipynb
            └── arasum_abstracts_detector.ipynb
    ├── Arabic_synthetic_dataset_generation/
        ├── AbstractsDataset/
            ├── allam.ipynb
            ├── analysis_on_the_generated_abstracts.ipynb
            ├── claude.ipynb
            ├── jais.ipynb
            ├── llama.ipynb
            ├── openai.ipynb
            └── top_frequent_words_analysis.ipynb
        ├── AraSum/
            ├── allam.ipynb
            ├── claude.ipynb
            ├── jais.ipynb
            ├── llama.ipynb
            └── openai.ipynb
        └── SocialMediaDataset/
            ├── allam.ipynb
            ├── analysis_on_the_generated_posts.ipynb
            ├── jais.ipynb
            ├── llama.ipynb
            ├── openai.ipynb
            └── top_frequent_words_analysis.ipynb
    └── exploration/
        ├── explore_arabic_content_detection_dataset.ipynb
        └── explore_arbicQA_dataset.ipynb
.gitattributes
.gitignore
LICENSE
README.md
requirements.txt

🔬 Research Methodology

Text Generation Strategies

Academic Abstracts (3 methods):

• Title-only generation: Free-form generation from paper titles
• Title+Content generation: Content-aware generation using paper content
• Abstract polishing: Refinement of existing human abstracts

Social Media Posts (1 method):

• Post polishing: Refinement preserving dialectal expressions

Models Evaluated

Model	Size	Focus	Source
ALLaM	7B	Arabic-focused	Open
Jais	70B	Arabic-focused	Open
Llama 3.1	70B	General	Open
GPT-4	-	General	Closed

Detection Approaches

• Binary detection: Human vs. Machine-generated
• Multi-class detection: Identify specific LLM
• Cross-model generalization: Train on one model, test on others

📊 Spotlight Findings

Stylometric Insights

• Reduced vocabulary diversity in AI-generated text
• Distinctive word frequency patterns with steeper drop-offs
• Model-specific linguistic signatures enabling identification
• Domain-specific characteristics varying between contexts

Detection Performance

Academic Abstracts:

• Binary detection: 99.5-99.9% F1-score
• Cross-model: 86.4-99.9% F1-score range
• Multi-class: 94.1-98.2% F1-score per model

Social Media:

• More challenging due to informal nature
• Cross-domain generalization issues confirmed
• Model-specific detectability variations observed

🚀 Getting Started

Prerequisites

# Python 3.8+
pip install -r requirements.txt

Installation

Make sure to also take a look at the corekit repo (https://github.com/MagedSaeed/llms-corekit) as it is needed in some scripts and notebooks.

git clone https://github.com/MagedSaeed/arabic-text-detection.git
cd arabic-text-detection

# Install dependencies
pip install -r requirements.txt

# Download datasets (requires Git LFS)
git lfs pull

The code was written with practices that support self-explanatory purpuses. You can browse the scripts and notebooks and run them providing the necessary API keys if required.

📁 Datasets

Academic Abstracts

• Source: Algerian Scientific Journals Platform
• Size: 8,388 samples across 3 generation methods
• Period: 2010-2022 (pre-AI era)
• Available: 🤗 HuggingFace Hub

Social Media Posts

• Source: BRAD (Book Reviews) + HARD (Hotel Reviews)
• Size: 3,318 samples (polishing method)
• Available: 🤗 HuggingFace Hub

🔗 Related Work

• Abstracts Data Collection code: arabic-dataset
• LLMs corekit: LLMs-corekit

📚 Citation

The Expert Systems with Applications Journal paper:

@article{al2025arabic,
  title={Arabic Machine-Generated Text Detection: Stylometric Analysis and Cross-Model Evaluation},
  author={Al-Shaibani, Maged S and Ahmed, Moataz},
  journal={Expert Systems with Applications},
  pages={130644},
  year={2025},
  publisher={Elsevier}
}

The preprint (arxiv):

@article{al2025arabic,
  title={The Arabic AI Fingerprint: Stylometric Analysis and Detection of Large Language Models Text},
  author={Al-Shaibani, Maged S and Ahmed, Moataz},
  journal={arXiv preprint arXiv:2505.23276},
  year={2025}
}

🏢 Institutional Support

This research is supported by:

• SDAIA-KFUPM Joint Research Center for Artificial Intelligence

👥 Authors

• Maged S. Al-Shaibani
• Moataz Ahmed - Corresponding Author (moataz.ahmed@kfupm.edu.sa)

SDAIA-KFUPM Joint Research Center for Artificial Intelligence
King Fahd University of Petroleum and Minerals, Saudi Arabia

⚖️ Ethical Considerations

This research is intended to:

• Improve detection of machine-generated content
• Enhance academic integrity in Arabic contexts
• Advance Arabic NLP research capabilities
• Support information verification systems

Please use this work responsibly and in accordance with ethical AI principles.

---