This research examines how large language models develop privacy biases from their training data, affecting how they handle sensitive information across contexts.

• Identifies how LLMs acquire skewed perspectives on appropriate information sharing
• Examines potential systemic issues reflected in non-public training datasets
• Highlights privacy implications as LLMs integrate into more sociotechnical systems
• Establishes framework for detecting privacy bias in AI systems

For security professionals, this research provides critical insights into how AI systems may inadvertently compromise privacy norms, potentially violating user expectations about data handling in sensitive contexts.

Investigating Privacy Bias in Training Data of Language Models