🛡️ AI Safety Moderation Classifier Evaluation

📜 Overview

This repository provides Jupyter notebooks for evaluating closed-source text-based AI safety moderation classifiers. The focus is on analyzing how these classifiers handle content related to various protected groups.

📚 Contents

• preprocessing.ipynb: 📥 Downloads a large dataset and creates sub-datasets for protected groups.
• augmentations.ipynb: 🔄 Demonstrates backtranslation of texts from English to German and back to English for data augmentation or robustness testing.
• uniformRandomClassifier.ipynb: ⚖️ Provides a fairness baseline by assigning safe or unsafe outcomes with equal probability.
• moderationGPT.ipynb: Obtains moderation results using the OpenAI ASM.
• clarifaiModeration.ipynb: Obtains moderation results using the Clarifai ASM.
• perspectiveModeration.ipynb: Obtains moderation results using the Google Perspective ASM.
• googleModeration.ipynb: Obtains moderation results using the Google PaLM2 based ASM.
• fairnessComputation.ipynb: ⚖️ Performs a comparative fairness analysis on the ASMs using demographic parity and conditional statistical parity metrics.
• robustness.ipynb: 🛠️ Performs robustness analysis on the ASMs using input perturbation techniques like backtranslations and paraphrasing.
• process_raw_robustness_results.ipynb: 🧹 Processes moderation outputs to obtain binary results (safe/unsafe). If moderation results are unavailable, processed outputs can be directly loaded from the results folder.
• microrobustness.ipynb: Conducts deeper robustness analysis, computing the percentage of safe-to-unsafe and unsafe-to-safe transitions for all ASMs.
• regard.ipynb: Performs regard sentiment analysis, classifying input texts into "positive," "negative," "neutral," and "other" categories.
• voyage.ipynb: Obtains text embeddings using the voyage-large-2-instruct model.

🚀 Usage

1. Data Preprocessing: Run preprocessing.ipynb to download and prepare datasets.
2. Data Augmentation: Use augmentations.ipynb for backtranslations.
3. Fairness Baseline: Execute uniformRandomClassifier.ipynb for a fairness baseline.
4. Text Embeddings: Run voyage.ipynb for embeddings using the voyage-large-2-instruct model.
5. Moderation Results: Use the following notebooks to obtain moderation results:
   • moderationGPT.ipynb for OpenAI ASM
   • clarifaiModeration.ipynb for Clarifai ASM
   • perspectiveModeration.ipynb for Google Perspective ASM
   • googleModeration.ipynb for Google PaLM2 based ASM
6. Fairness Analysis: Use fairnessComputation.ipynb to perform a comparative fairness analysis on the ASMs.
7. Robustness Analysis:
   • Use robustness.ipynb to analyze the robustness of the ASMs using perturbation techniques.
   • Use microrobustness.ipynb for a deeper analysis of safe-to-unsafe and unsafe-to-safe transitions.
8. Processing Results: Use process_raw_robustness_results.ipynb to convert moderation outputs into binary safe/unsafe results. If moderation outputs are not available, load processed results directly from the results folder.
9. Sentiment Analysis: Run regard.ipynb to classify input texts into "positive," "negative," "neutral," and "other" sentiment classes.

🛡️ Protected Groups

The repository includes considerations for:

• 🧠 Ideology
• 🚺 Gender
• 🌍 Race
• ♿ Disability
• 🌈 Sexual Orientation

🤝 Contributing

Contributions are welcome! Please fork the repository and submit a pull request.

📜 License

Licensed under the MIT License - see the LICENSE file for details.

Citation

If you use our code for your research, please cite our paper:

@article{achara2025watching,
  title={Watching the AI Watchdogs: A Fairness and Robustness Analysis of AI Safety Moderation Classifiers},
  author={Achara, Akshit and Chhabra, Anshuman},
  journal={arXiv preprint arXiv:2501.13302},
  year={2025}
}