How AI Models Learn Bias from Synthetic Training Data

AI Ethics with Fexingo: Bias, Safety, and Responsible Artificial Intelligence · Jul 25, 2026 · 5m

By 2026, an estimated 60% of training data for large language models will be synthetic — AI-generated text used to teach other AIs. But a new MIT Media Lab study shows that synthetic data retains about 80% of the gender and racial biases from its original model. Meanwhile, Oxford researchers found that successive generations of training on synthetic data can lead to 'model collapse': outputs become less creative, more repetitive, and more stereotypical after just three generations. This isn't just an academic problem. A customer service platform might train its chatbot on synthetic…