Privacy risks from medical AI tools are not shared equally

TL;DR - Privacy attacks can expose whether an individual’s medical data was used to train an AI model, with people who differ from the majority facing greater risk. This highlights unequal privacy harms in medical AI.

• Training-data membership can potentially be inferred through privacy attacks.
• Medical records from underrepresented or atypical individuals are especially vulnerable.
• Aggregate privacy assessments might obscure disparities between demographic or clinical groups.