Language and ageist stereotypes in generative conversational AI systems in Spanish

Abstract

This article explores linguistic ageism in Spanish in Generative Artificial Intelligence (GAI) systems, examining how these technologies reproduce or challenge discriminatory practices towards older adults in discourse. The study was conducted on six conversational GAI models (ChatGPT, YouChat, Copilot, Perplexity, Gemini, and DeepSeek). The research was structured in three phases. In the first phase, a Likert-type scale with 48 statements was applied to each model, with the aim of quantitatively evaluating the discriminatory language internalised by each system. In the second phase, the IAGs were asked to generate texts on ten dimensions related to older people to find discursive patterns and prototypical positive and negative stereotypes. Finally, a content analysis was carried out to identify lexical and structural patterns in the texts generated. The results show that DeepSeek and ChatGPT were the most effective models in identifying ageist expressions, while Copilot showed lower levels of detection. Common difficulties were identified in detecting positive ageism. At the discursive level, all models reproduced ageist stereotypes, although Gemini and DeepSeek offered more balanced and positive representations of old age. Finally, it was found that the syntactic structures and statistical significance of the language produced by AGIs represent older people as passive and dependent subjects. These findings suggest that IAGs reflect and perpetuate discriminatory ageist patterns, making it necessary to review them so that they can act as active tools in promoting neutral and inclusive language in Spanish.