Large language models are built on text corpora that overwhelmingly reflect Western, anglophone perspectives. According to a new arXiv paper, this asymmetry is not neutral: it shapes how models respond to questions about territory, community authority, and the relationship between people and land. The result is that LLMs can implicitly reinforce colonial assumptions even when answering seemingly factual queries.
The paper suggests Constitutional AI as a corrective. This approach involves defining a set of principles—drawn here from Indigenous cosmovision—that the model is explicitly trained and constrained to follow. Rather than treating Western norms as a default, the model would be aligned to respect Indigenous ways of knowing and relating to the world.
The proposal is significant because it moves beyond generic "debiasing" toward a positive framework for embedding specific cultural values into AI systems. The abstract does not report experimental results, so the practical effectiveness of the approach remains to be seen. But it offers a concrete path for making LLMs more pluralistic in how they understand authority, community, and place.