PhilMod Talk - AI Model Collapse and our Epistemic Ecosystem (Jovy Chan)

PhilMod Talk - AI Model Collapse and our Epistemic Ecosystem (Jovy Chan)

Join us for a discussion with Jovy Chan, Postdoctoral Ethics Fellow at Stanford University's McCoy Center for Ethics in Society. Presentation (30 min.) + discussion (30 min.) More information about our community and speaker series is available at philmod.org

Abstract: More people are now using LLMs in their day-to-day tasks. More content on the internet is now AI-generated. This trend prompted AI developers to worry about a problem called Model Collapse. Model Collapse arises because AI models are trained on data pulled from the internet, and when the internet is filled with content generated from the AI models themselves, the performance of newer models will deteriorate instead of improve. Model collapse is a serious problem, but not only for AI developers. It is also a warning alarm that our epistemic ecosystem is undergoing a fundamental shift. In particular, it signals that the widespread use of LLMs will drastically change the online repository of public data, data that we have come to rely on heavily for the exchange of knowledge. This talk will explore what exactly these changes are, why they are problematic, and what we can do to mitigate their effects.