Colloquia

Samenvatting

This thesis explores the challenge of integrating safety standards into European education, despite their vital role in creating a safe and reliable society. A key problem identified is that current safety standards are poorly incorporated as learning materials for a diverse student population. Students often find them hard to access, understand, or apply due to differences in literacy, background knowledge, and the complexity of the documents themselves.

Existing solutions, such as e-learning platforms, professional courses, and gamified approaches, have limitations, particularly in offering personalized and context-relevant support. While generative AI shows potential to bridge this gap, generic models often lack the specialized knowledge necessary for safety instruction.

To address this, the thesis introduces SIAI (Safety Instructor AI), a system of customized GPT-based chatbots grounded in content from the Safety by Design course. SIAI offers tailored, textual and visual support to help students better navigate, understand, and apply the NEN NTA 8287:2021 safety standard.

The system was evaluated through three tests:

• Robustness: SIAI delivered both general and domain-specific responses with high reliability. It achieved 100% accuracy and completeness in short-form answers, and effectively interpreted varied communication styles.
• Visual Comprehension: Using visual mock-ups, the chatbot’s hierarchical structure was found to improve clarity and engagement, proving more effective than traditional layouts.
• Comparative Study: In a task-based experiment, students using SIAI outperformed those using standard documents and Google Search in retrieving information, understanding terminology, and applying safety steps in real-world design tasks.

The findings show that generative AI, when tailored to a specific domain, can go beyond general assistance: it can meaningfully enhance how complex standards are explained.

However, while the prototype exceeded expectations in small-scale testing, the discussion notes that further large-scale validation is essential. Questions remain about long-term learning outcomes, integration into curricula, and maintaining accuracy as standards evolve“.