The Geometry of Belonging: The Euclidean Error in Large Language Model Processing of Anishinaabe Knowledge Systems
Publication information:
Abstract
Standard LLM architectures assume Euclidean geometry: semantic proximity measured in flat vector space. Anishinaabe knowledge systems do not organize meaning this way. Kinship in Anishinaabemowin is non-transitive, cyclic, and multi-scalar, encoding the belonging of ancestors, land, and spiritual persons as grammatical structure. Euclidean distance does not approximate these structures poorly. It destroys them. The paradox: the more Anishinaabemowin data LLMs ingest, the more thoroughly its relational architecture is overwritten. The language arrives; the knowledge structure does not. I name this the Euclidean Error, propose Topological Data Analysis as its structural remedy, and report results for an Indigenous-governed language revitalization game.
Presenter Biography
Orus Mateo Castaño-Suárez, M.A., is accepted to the Digital Media PhD program at York University’s Lassonde School of Engineering and is an Ontario Graduate Scholar. In Abundant Intelligences, they co-create Indigenous AI sovereignty research across language models, archives, AR placekeeping, EEG-informed dream-symbol analysis, and topological knowledge mapping. OrusMateo.com