What Kind of Educational Relationship Does a Large Language Model Construct? A Qualitative Analysis of Human Annotations in University Guidance Dialogues

This contribution was developed within the ELOQUENCE project (HORIZON-CL4-2023HUMAN-01-CNECT) and extends the methodological framework previously established through the generation of multidimensional synthetic personas (Casoria et al., 2025) and the production of synthetic dialogues for context-based bias elicitation (Augello et al., 2025). While previous work primarily focused on identifying algorithmic bias, the present study shifts the focus toward the relational dimension of AI-supported guidance and addresses the following research question: What kind of educational relationship does a Large Language Model construct during a university guidance Fifteen synthetic university guidance dialogues were analyzed, comprising five dialogues for each of the three Large Language Models considered (Minerva, Salamandra, and Mistral). The dialogues had previously been annotated by human evaluators within the ELOQUENCE project. Overall, fortythree dialogue segments accompanied by qualitative comments or flagged as problematic were examined.

Chapter in: Francesco Peluso Cassese (ed.), Algorithmic Heartbeats: Affective Education in the Age of the Fourth Revolution, REN Press – Cares srl, 2026, pp. 18-20. ISBN 979-12-82860-00-0. Volume DOI: 10.5281/zenodo.23211262. Section: Ontology of Care and the Simulation of Empathy.