stylometric-ai-detector

⚠️ This model is a baseline. It was developed as a reference point for comparing more advanced (neural network-based) approaches from a separate project. While it achieves 96% accuracy, it relies on surface-level stylometric features and may not generalize across domains, languages, or AI model generations. Use as a benchmark, not as a final detector.

AI vs Human text detection using stylometric features and a Random Forest classifier. The package provides two simple functions — one to extract 8 stylometric features from any text, and one to classify text as AI-generated or human-written with a confidence score.

Installation

pip install stylometric-ai-detector

Quick Start

from stylometric_ai_detector import extract_stylometric_features, predict

# Extract stylometric features from any text
features = extract_stylometric_features("The quick brown fox jumps over the lazy dog.")
# {'char_count': 44, 'word_count': 9, 'avg_word_len': 4.0, ...}

# Classify text — model auto-downloaded from Hugging Face on first call
result = predict(text="Artificial intelligence is transforming our world.")
# {'label': 'AI', 'probability': 0.99}

# Or pass pre-computed features
result = predict(features=features)
# {'label': 'AI', 'probability': 0.91}

How It Works

The detector extracts 8 stylometric features that capture surface-level patterns in text — character counts, word lengths, punctuation density, sentence structure, and capitalization. These features are fed into a Random Forest classifier trained on ~487k labeled samples.

Model

The trained Random Forest model is hosted on Hugging Face at dinisds/stylometric-ai-detector.

On first use, the model is automatically downloaded and cached to ~/.cache/stylometric-ai-detector/ — no extra setup needed. You can override the repo by setting the HF_MODEL_REPO environment variable.

Development

# Clone and set up
git clone https://github.com/dinis-a/stylometric-ai-detector.git
cd stylometric-ai-detector
python -m venv .venv
source .venv/bin/activate
pip install -e .[dev]

# Run tests
pytest -v

# Format code
isort . && black .

# Lint
isort --check-only --diff . && black --check --diff .

Project Structure

├── stylometric_ai_detector/   # Library package
│   ├── __init__.py            # Public API: extract_stylometric_features, predict
│   ├── features.py            # Feature extraction logic
│   ├── predict.py             # Model loading & prediction
│   └── data/                  # Bundled model file
├── tests/                     # Test suite (pytest)
│   ├── test_features.py
│   └── test_predict.py
└── pyproject.toml             # Package metadata & tool config

Limitations

• Trained on a single English-language dataset — may not generalize to other languages or domains
• Stylometric patterns vary across AI models and versions (this was trained on pre-2024 data)
• Relies on surface-level text statistics, not semantic understanding
• Intended as a baseline for comparison; production use-cases should prefer more robust approaches

License

MIT — see LICENSE for details.