Mapping Generative AI

Training data is foundational to generative AI systems. From Common Crawl's 3.1 billion web pages to LAION-5B's corpus of almost 6 billion image-text pairs, these vast collections - scraped from the internet and treated as "ground truth" - play a critical role in shaping the epistemic boundaries that govern generative AI models. Yet training data is beset with complex social, political, and epistemological challenges. What happens when data is stripped of context, meaning, and provenance? How does training data limit what and how machine learning systems interpret the world? What are the copyright implications of these datasets? And most importantly, what forms of power do these approaches enhance and enable? This keynote is an invitation to reflect on the epistemic foundations of generative AI, and to consider the wide-ranging impacts of the current generative turn.