Conceptual framework for evaluating cognitive and discursive characteristics of LLMs’ narratives in environmental science

The application of large language models (LLMs) for research and scientific discourse is expanding because of their use in generating narratives, explaining complex concepts, and influencing human perception of various issues, including climate change and sustainability. However, a systematic approach for evaluating the quality and cognitive effects of LLMs’ output remains underdeveloped. This study proposes a conceptual framework for the evaluation of narratives generated by LLMs for the cognitive and discursive characterization of texts in a structured, comparable manner. The study integrates the six dimensions: cognitive framing, information selection/exclusion, emotional modulation, cognitive load and accessibility, trustworthiness/epistemic positioning, and persuasive potential. Each dimension reflects how LLMs construct the meaning of narratives and shape users’ perception. Each dimension can be assessed using diverse analytical strategies and openly available tools, such as lexical analysis platforms (e.g., Voyant), readability indices, sentiment analysis, stance detection, and rhetorical analysis. This study may serve as an example of establishing a precise, repeatable framework for comparing the LLMs-generated narratives in English and to enhance the responsible use of generative AI in environmental discussions by standardizing evaluation methods and linking cognitive-discursive analysis with sustainability communication.