Aim. The present research aims to identify the mechanisms of processing intensional linguistic deviations by large language models (LLMs) and to analyse their linguocreative strategies in digital discourse. The aim of the work is to develop a theoretical model explaining cognitive algorithms for recognising and transforming deviations in artificial intelligence systems. Methodology. An integrated approach including corpus analysis of a diachronic slice of digital discourse (2019–2024), experimental prompts with controlled deviations for GPT-4, Gemini 1.5 and Claude 3 models, and discourse analysis of AI speech acts using a three-level annotation scale (replication/amplification/normalisation) was used as a methodological framework. Results. The results of the research supported the hypothesis of the statistical nature of LLM linguocreativity, revealing a three-stage model of deviation processing: recognition through attention mechanisms, classification by degree of deviation from the norm, and strategic choice of response. The paradox of ‘creative conformism’ is established, manifested in AI's tendency to hypernormalise initially marginal linguistic innovations. Of particular practical interest are the documented effects of circulation of AI-generated neologisms in social media and the formation of ‘artificial linguistic taste’. Research implications. The theoretical significance of the work lies in the development of the apparatus of cognitive linguistics of digital discourse and clarification of the ontology of intensional deviations. Practical value is related to applications in the field of NLP-systems development, digital linguodidactics and prediction of language changes. The data obtained opens prospects for further studying culturally specific deviations in multilingual models and developing metrics for assessing the linguocreative potential of AI.