Fraseologismo sem mundo: expressões idiomáticas não convencionalizadas, inteligência artificial generativa e letramento fraseológico crítico

This article analyzes how two generative artificial intelligence models, ChatGPT and Microsoft Copilot, accept non-conventionalized lexical combinations as idiomatic expressions—both in Brazilian Portuguese and in English—created by analogy with authentic units from these languages. Drawing on Freire’s notion of reading the world, the research relates human phraseological competence, simulated artificial phraseological competence, the concept of the classeme, and the proposal of a critical phraseological literacy. Seventy-nine acceptances, drawn from a previous study with 360 responses, were analyzed through independent coding by two language models, with constant review by a human researcher, adopting a qualitative approach at the classematic, phraseological, and literacy levels. The results indicate that the models either assign the meaning of the base unit or build one from compositionality, inventing etymologies, sociopragmatic registers, regional variations, and even proposals for didactic application, which gives a fabricated and artificial world to these worldless sequences. It is argued that critical phraseological literacy enables teachers and learners of a first or additional language to question this authority by verifying the authenticity of these units in independent sources before applying them to teaching.