How Synthetic Data Bakes In Hidden Bias

AI Ethics with Fexingo: Bias, Safety, and Responsible Artificial Intelligence · May 22, 2026 · 9m · transcript

Episode 5 of AI Ethics with Fexingo tackles a next-generation bias problem: training data generated by AI itself. Lucas explains how synthetic data — used to save costs and protect privacy — can amplify existing prejudices when models learn from their own outputs. The conversation centers on one concrete case: a medical imaging model that inherited skin-tone bias from a synthetic dataset derived from a real dataset that already underrepresented darker skin. Lucas and Luna walk through the feedback loop, why it's hard to detect, and what regulators like the EU AI Act are starting to require. A…