Machine-In-The-Loops. Reclaiming Human-Centered Knowledge Bulding and Decision-Making in AI for Education

This thesis addresses the question of how to realise the full potential of Artificial Intelligence in Education (AIED), identifying three forms of uncertainty that hinder its responsible adoption: technological (the opacity of models and the problem of explainability), socio-technical (the invisibility of the interests and actors that shape these systems), and educational (the still poorly understood effects on the learning and autonomy of students and teachers). It argues that these uncertainties are not independent, but constitute a recursive causal system in which each reinforces the others. In response, the thesis proposes the Machine-in-the-Loop (MITL) paradigm, which places human expertise back at the centre of the epistemic and decision-making cycles of AIED through three principles: epistemic precedence, causal transparency, and intervention orientation. The framework is illustrated through two case studies applied to the prediction of low achievement (INVALSI dataset) and the prevention of university dropout.