Demographic Background Prompting Does Not Affect Linguistic Features on LLM-Generated News Texts

• Alberto Muñoz-Ortiz ; Carlos Gómez-Rodríguez ^[1] ; David Vilares ^[1]
  1. [1] Universidade da Coruña

     Universidade da Coruña

     A Coruña, España

     
• Localización: Proceedings XoveTIC 2024: Impulsando el talento científico / coord. por Manuel Lagos Rodríguez, Tirso Varela Rodeiro, Javier Pereira-Loureiro, Manuel Penedo, 2024, págs. 169-176
• Idioma: inglés
• Enlaces
  • Texto completo
• Resumen

  • We explored if implicit demographic information in prompts for large language models (LLMs) influences the linguistic features of generated text. Two LLMs were prompted to write news articles based on a title and summary, with prompts including demographic details like age, income, or nationality. The models were instructed not to explicitly reference these details. A total of 28,080 articles were generated by varying the demographics and topics. We calculated various linguistic metrics (e.g., sentence length, type-token ratio) and performed ANOVA, treating linguistic metrics as dependent variables and demographic categories as independent variables. Results indicate that demographic attributes do not significantly impact the linguistic metrics.