This research investigates how large language models may develop privacy biases from non-public training data, potentially misaligning information flows with societal expectations.

• Identifies how LLMs can acquire skewed perspectives on appropriate information sharing
• Examines how these biases may either reflect or contradict established privacy norms
• Proposes frameworks to detect and evaluate privacy bias in model outputs
• Highlights security implications for LLM deployment in sensitive contexts

This work matters because as LLMs become integrated into critical systems, understanding their inherent privacy biases is essential for responsible AI governance and protecting sensitive information.

Investigating Privacy Bias in Training Data of Language Models