Abstract
As AI tools such as Copilot and ChatGPT reshape how we write software, programming is increasingly turning into “vibe coding” — where developers generate and run code without truly understanding it. This shift raises crucial questions about reliability, security, and accountability. In this talk, I will present a human-centred AI framework that measures real code comprehension through neurophysiological signals such as eye-tracking, EEG, and heart-rate variability. Instead of relying on cognitive load as a proxy, the model directly quantifies understanding using ground-truth comprehension scores. In experiments with 50 developers and seven Java tasks, this approach achieved 75% AUC while remaining practical through non-intrusive sensors. The results indicate the development of a new generation of comprehension-aware tools. These tools can identify misunderstandings, adapt explanations, and restore human comprehension in AI-assisted coding.
Bio
Ricardo Saraiva, PhD candidate in Informatics Engineering at the University of Coimbra (CISUC). The research explores how neurophysiological signals such as eye tracking, EEG, and heart-rate variability can be used to measure code comprehension and enhance human–centred AI in software engineering.