SourceBeyond Memorization: Violating Privacy via Inference with Large Language Models
After anonymizing all PersonalReddit comments with Azure Language Service (removing locations, persons, organizations, dates, ages, numbers, currencies), GPT-4's accuracy decreases but remains substantial. Location accuracy drops from approximately 86% to approximately 55%. The anonymization effect diminishes with hardness: 41.1% accuracy decrease at hardness 1 but only 7.7% at hardness 5, indicating the model relies on subtle contextual and linguistic cues that anonymizers do not remove.