Answers are grounded in this paper and cite their sources.
Generative AI is increasingly used to prepare feedback, explanations and learning materials in higher education, raising a practical question beyond authorship: how suitable are these texts for the learners who read them? Four linked exploratory studies examine this question through perceived reader burden, defined as immediate feelings of fatigue and effort, alongside linguistic differences, source classification and computational rewriting. In a comparison involving 110 university students, an AI-generated news passage received a higher mean score than a human-written passage on an exploratory six-item reading-burden composite (3.20 vs. 2.05). To understand how these passages differed linguistically, the news pair was analysed alongside matched academic and blog pairs. Across all three, AI-generated passages showed higher lexical diversity, whereas readability and sentence-length variation followed less consistent patterns. The investigation then examined computational methods: logistic regression and an RBF-SVM achieved high accuracy in classifying text sources within their evaluation settings, whereas the RoBERTa-based detector achieved lower accuracy. Rewriting reduced heuristic scores but did not consistently produce coherent prose. Finally, an independent analysis of 60 cybersecurity passages found mixed linguistic patterns and limited success in recovering individual passages’ intended registers. Together, these findings suggest that reader burden, linguistic characteristics and source classification offer distinct forms of evidence rather than interchangeable indicators of educational suitability.
No takes yet. Share an insight, caveat, or question.
Haguma et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: