When AI Trains on AI: The Model Collapse Problem

Logosofia no Dia a Dia · Jul 12, 2026 · 28m · transcript

By mid-2026, over 60% of web text is AI-generated. When new AI models train on that synthetic content instead of human data, a dangerous feedback loop emerges. This episode explores model collapse — the technical phenomenon where AI systems degrade after just five generations of training on their own outputs, losing rare knowledge, flattening language, and erasing the long tail of human experience. We trace the concrete consequences: from medical AIs that miss rare diseases to a homogenized internet where every search result says the same thing. A deep dive into what happens when the snake…