Directional Semantic Drift Across Grokipedia Versions: A Diffusion-Manifold Analysis

Large language models increasingly mediate access to factual knowledge, yet little is known about how their representations diverge from established encyclopedic sources. We quantify semantic drift between Wikipedia and Grokipedia–a large language model– generated encyclopedia explicitly designed to counter perceived Wikipedia biases–as well as across two Grokipedia versions released before and after a platform update (v0.1, v0.2). Focusing on articles about cities in the United States and Germany, we embed 1,387 matched entries, project them onto a shared diffusion manifold, and operationalize semantic drift as Euclidean displacement in latent space. Both cross-sectional and longitudinal analyses reveal systematic structure in this drift: larger cities and articles containing politically salient framing exhibit significantly greater divergence. Regression models further show that changes in article length and political content predict drift even after controlling for population size, country, and time. Articles in Grokipedia v0.1 that were explicitly attributed as adapted from Wikipedia lost this attribution in v0.2 and exhibited a pronounced and statistically significant semantic shift, indicating a new article generation. ... mehrThe proposed diffusion-manifold framework provides a scalable testbed for researching semantic bias and temporal drift in AI-mediated knowledge.