Gael Varoquaux • 2/13/2025

AI for health: the impossible necessity of unbiased data

This article examines the necessity and impossibility of unbiased data for building AI in healthcare. It discusses how historical choices and population sampling create bias, leading to discrimination (e.g., pulse oximeters on dark skin). It argues that 'unbiased' is a value-laden concept and that AI models must account for causal effects and comparable populations to make good medical decisions.

0 comments

#Machine Learning #ai ethics #Healthcare