Large language models (LLMs) are now widely used in consumer-facing applications worldwide. However, the majority of empirical research on how individuals develop trust in LLM-generated content has primarily focused on Western, English-speaking, and highly connected populations. To address this gap, we present the first large-scale, controlled empirical study examining trust perceptions of LLMs in African contexts. We recruited 412 participants from 12 African countries across West, East, Southern, and Central Africa, who assessed African-contextualised factual texts in three domains (news, science, and legal) that varied by disclosed source (human vs LLM) and disclosure condition (disclosed vs blind). We measured trust using three subscales: competence, integrity, and benevolence, alongside perceived credibility, information quality, and the intention to fact-check. Additionally, we captured African-specific variables, including type of internet access, urban or rural classification, primary language, and news consumption platform, as individual-difference covariates. Our findings indicate a significant penalty for source disclosure. LLM-labelled texts received an average trust score that was 34.3% lower (M = 3.08) compared to human-labelled texts (M = 4.63). Notably, participants were unable to distinguish LLM-generated text from human-generated text at a rate better than chance (53.8%, p = .13) when labels were absent. Furthermore, AI literacy was found to significantly moderate this penalty.

PDF