This paper presents the trustworthiness assessment of an AI-driven platform for inquiry-based STEM education,
deployed in real secondary school classrooms. The assessment was conducted using a structured trustworthiness
risk assessment methodology, covering cartography, threat analysis, consequence assessment, vulnerability
analysis, risk analysis, and risk management, and spanning all AI lifecycle phases—design, development, and
deployment. Following this methodology, a set of mitigation controls was identified and implemented across the
platform’s three core AI models: the Learning Companion, the Student Performance Classifier, and the Learning
Recommendations system. After the implementation of these controls, the models achieved the following results:
the Learning Companion achieved 100% intent classification accuracy across all intents; the Student Performance
Classifier reached 96.7% accuracy on unseen data; and the Learning Recommendations system attained an 80%
teacher acceptance rate in its initial deployment. The assessment will be repeated in subsequent pilot phases.
These results illustrate the application of the trustworthiness assessment methodology in a real-world educational
context, providing a replicable methodology for EU AI Act-compliant AI deployment in high-risk educational
settings.
