BCAM, October 23th 2026, 13:30 - 14:30
When models eats their own words: model collapse.
Francisco Javier Sanguino Bautiste (BCAM).
With the increasing use of large language models (LLMs), people are filling the internet with AI-generated content. This raises an important question: what happens when AI learns from AI? This talk introduces model collapse: a process in which repeated training on generated data poses a major risk—diversity can shrink and rare patterns can gradually disappear. Yet much of the literature leaves human agency outside the picture. People select, edit, and reshape model outputs before sharing them online, influencing what future models learn. We will explore whether human intervention is all you need, and whether removing humans from the loop is THE problem. If so, self-improving AI may face greater technical challenges than Elon, Sam, and Dario would have us believe.