Artificial intelligence (AI)-generated synthetic data has emerged as a promising solution to address the underrepresentation of underserved populations in medical AI systems. By artificially generating data that mimics real-world patient information, proponents argue that AI-generated synthetic data can fill data gaps, improve algorithmic fairness, and mitigate bias without requiring costly data collection or raising privacy concerns. However, in this article, I challenge this optimistic view. I argue that AI-generated synthetic data may amplify existing biases rather than mitigate them, and complicate informed consent processes in ways that disproportionately harm underserved populations. I examine two critical ethical dimensions: (1) how the opacity inherent in synthetic data generation can reproduce and amplify discriminatory patterns present in the source data, and (2) how AI-generated synthetic data complexifies informed consent. I conclude that while AI-generated synthetic data offers apparent technical advantages, it fails to address, and may worsen, data disparities affecting underserved populations. Rather than pursuing synthetic solutions, I believe that we should address the structural barriers that prevent genuine inclusion of underserved populations in medical research.
Stéphanie Baggio (Tue,) studied this question.