Mapping AI feedback under emotional prompts: a sentiment-topic-network approach

Purpose – While generative artificial intelligence (AI) is increasingly used to generate feedback on student writing, little is known about how emotional prompts affect the structure of such feedback. This study aims to examine how large language models respond to positive, neutral and negative prompts in postgraduate scientific writing, focusing on how emotional cues shape sentiment polarity, thematic content and structural composition in AI-generated feedback and how these patterns inform learning-relevant feedback architectures. Design/methodology/approach – An integrated sentiment-topic-network framework of VADER sentiment analysis, Latent Dirichlet allocation topic modelling and epistemic network analysis (ENA) was used to analyse 198 AI-generated feedback messages on 66 postgraduate scientific reports. Sentiment analysis classified emotional polarity, topic modelling identified thematic clusters within which sentiments were expressed, and ENA modelled co-occurrence patterns among cognitive, affective and dialogic codes to characterize feedback architectures under positive, neutral, and negative prompts. Findings – Across all conditions, positive sentiment dominated even when prompts were neutral or negative, with only negative prompts causing substantial increases in negative codes. Topic–sentiment patterns showed that certain feedback themes were selectively intensified by emotional tone. ENA identified three distinct feedback architectures: supportive-integrative networks under positive prompts, fragmented–minimal networks under neutral prompts, and critical–precision networks under negative prompts. Together, these patterns indicate that emotions modulate not only what feedback is generated but also how cognitive and affective elements are integrated. Research limitations/implications – Data were drawn from 66 postgraduate students at a single institution, yielding a small, context-specific convenience sample. Findings are therefore interpreted as exploratory patterns rather than statistically generalizable estimates, with analytic generalization aimed at comparable science, technology, engineering and mathematics (STEM)-oriented postgraduate contexts. We did not measure achievement gains or long-term retention. Future research should examine more complex emotional contexts and prompt designs, use mixed-method approaches that include qualitative analysis of emotional nuance and empirically test the three hypotheses about revision quality, perceived usefulness, trust and emotional safety. Practical implications – Educators can embed prompt templates corresponding to supportive–integrative, fragmented–minimal and critical–precision architectures into course guidelines and teach students when and how to invoke them. Positive prompts can be used to normalize difficulty and sustain engagement in early drafting, with brief episodes of critical–precision feedback introduced once drafts are coherent. Developers of institutional AI tools can implement visible “feedback mode” selectors that map onto these architectures, allowing teachers to align system behaviour with course-level pedagogy rather than relying on a single default. Originality/value – This study advances an integrated Sentiment–Topic–Network framework that links sentiment polarity, thematic focus and network-level co-occurrence patterns in AI-generated feedback. It empirically characterizes a robust positivity bias in large language model feedback under positive, neutral, and negative prompts and identifies three feedback architectures – supportive-integrative, fragmented–minimal and critical–precision – relevant for emotionally tuned feedback design. By translating these architectures into an adaptive model and testable hypotheses, the study offers a reusable lens for analysing emotionally primed AI feedback in postgraduate STEM-oriented learning contexts.