How AI Learning Models Are Trapped by Their Own Data

AI Ethics with Fexingo: Bias, Safety, and Responsible Artificial Intelligence · Jul 21, 2026 · 13m

In this episode of AI Ethics with Fexingo, Lucas and Luna explore a surprising source of bias: when AI models are trained on data generated by older AI systems. They dive into a 2025 study from researchers at Oxford and MIT showing that language models fed synthetic text from earlier models develop 'model collapse'—becoming less diverse and more error-prone over generations. Lucas explains why this matters for everything from customer service chatbots to medical diagnostic tools, and Luna pushes him on whether this is a fixable technical problem or a fundamental limit. They also discuss the…