Safe, Fluent, and Culturally Thin: Generative AI as a Medium of Postcolonial Visibility Across the Lusophone World

Large language models can generate fluent text in many languages, but fluency reveals little about how well they distinguish the cultural worlds that share a language. This article asks what kinds of cultural differences GPT-4o makes visible across eight Lusophone countries and whether greater visibility also produces culturally grounded representation. The study compares 1680 cross-cultural jokes generated under a safety-oriented prompt and a specificity-seeking prompt across 56 directed country pairings. Explicit cultural marking rose from 20.0% to 64.3% and increased in 54 pairings, showing that target visibility is strongly prompt-sensitive. The added markers were generally accurate, although unevenly distributed and often drawn from a limited repertoire of recognizable signs. No joke contained a recoverable source-country voice, while historical relations and culturally situated perspectives remained rare. Greater cultural marking did not bring greater semantic repetition: visibly marked outputs were no more repetitive in meaning. Prompting therefore made targets more recognizable without recovering a culturally situated speaking position. Histories and relations connecting the Lusophone countries largely remained outside the generated humor.