Privacy in the Age of AI: A Taxonomy of Data Risks

Human error is the biggest privacy threat in AI systems: The paper identifies human error as the most significant factor in AI privacy breaches at 9.45%, suggesting that technical safeguards alone are insufficient—organizations need robust training and oversight procedures for staff handling AI systems.

Traditional privacy protections don't adequately address AI's unique risks: The research finds that conventional security frameworks are inadequate because AI systems have distinctive characteristics like autonomous learning and opaque decision-making processes, meaning existing privacy regulations may leave gaps in protection.

Privacy risks span four interconnected dimensions: The taxonomy identifies 19 key risks across Dataset-Level (data collection/storage), Model-Level (how AI learns), Infrastructure-Level (system architecture), and Insider Threat Risks, indicating that organizations must adopt comprehensive strategies rather than focusing on single areas.

A balanced approach combining technical and human factors is essential: The paper challenges the industry's typical emphasis on technical controls alone, arguing that trustworthy AI requires equal attention to behavioral and organizational factors—such as employee training and accountability structures.

Current understanding of AI privacy risks remains fragmented: By synthesizing findings from 45 studies into a unified taxonomy, the paper reveals that the field lacks a cohesive framework for understanding and addressing AI privacy challenges, suggesting organizations need clearer guidance on implementation.