Linguistic Relativity Meets Large Language Models: Does Prompting in Yorùbá Elicit Qualitatively Different AI Outputs Than Prompting in English?

Abstract

The Sapir-Whorf hypothesis of linguistic relativity posits that the language one uses shapes the nature of one's thought. With large language models (LLMs) now serving as generative intermediaries across dozens of languages, a critical empirical question arises: does the language of a prompt qualitatively change the structure, cultural grounding, and epistemic character of LLM outputs? This paper presents the first controlled bilingual prompting experiment comparing English and Yorùbá inputs on a domain-specific rural electrification corpus, grounded in a real AI literacy training programme delivered to the Oyo State Rural Electrification Board (OYSREB) by Rollascriptings MastersCryptosLLMs. Using a novel Semantic Grounding Index (SGI) and Abstractness Ratio (AR) framework, we show that Yorùbá-prompted outputs exhibit a statistically significant 521% increase in cultural grounding (mean SGI: 0.795 vs 0.128, t(9)=−8.41, p<0.001, Cohen's d=−3.999) and a 61.9% reduction in abstractness (mean AR: 0.287 vs 0.761, t(9)=13.13, p<0.001, Cohen's d=6.846), while maintaining equivalent lexical output length. These results constitute empirical support for a soft linguistic relativity effect in LLMs: language does not merely translate content but restructures the epistemic character of AI-generated knowledge. We discuss implications for multilingual AI deployment in low-resource African language contexts.