Extracting PAC Decision Trees from Black Box Binary Classifiers: The Gender Bias Case Study on BERT-based Language Models

Decision trees can help us understand "black box" AI models: The paper demonstrates that simpler, transparent decision trees can be extracted from complex language models like BERT to reveal how these models make decisions, making their behavior more interpretable to non-experts.

We can now mathematically guarantee how well explanations match the original model: Using the PAC (Probably Approximately Correct) framework, researchers can provide theoretical guarantees about how faithfully an extracted decision tree represents the original model's behavior, addressing a critical trust issue in AI explainability.

BERT language models exhibit occupational gender bias: The case study found that BERT-based models show systematic gender bias in occupational contexts, meaning these widely-used language models may make gender-biased predictions about job-related tasks.

This method provides a practical tool for auditing AI systems: By extracting interpretable decision trees with mathematical guarantees, organizations can systematically audit commercial AI models for biases and other problematic behaviors without needing access to the model's internal workings.

Explainability and bias detection are interconnected: The paper shows that making models more explainable through decision tree extraction simultaneously enables researchers to identify and document specific biases, suggesting that transparency efforts serve dual purposes in responsible AI development.