This research reveals how manipulating personal information (PII) in training data creates broader privacy effects than previously understood.

• Ripple Effects: Changes to one person's data influence how LLMs memorize information about other individuals
• Correlation Impact: Individuals with data similar to added/removed PII experience heightened privacy risks
• Dynamic Memorization: PII memorization is not static but fluctuates based on training data composition
• Security Implications: These findings challenge current approaches to data removal requests and privacy compliance

For security professionals, this research demands rethinking privacy protection strategies beyond individual opt-outs, suggesting more comprehensive approaches to managing personal data in AI systems.

Privacy Ripple Effects from Adding or Removing Personal Information in Language Model Training