Explainability and Certification of AI-Generated Educational Assessments

A team of researchers has published a chapter outlining a critical framework to bring trust and accountability to AI-generated educational tests. The work, led by Antoun Yaacoub, Zainab Assaghir, and Anuradha Kar, addresses a major barrier to adoption: the current lack of transparent, explainable, and certifiable mechanisms for AI-created assessment items like quizzes and exams. Their proposed system tackles this by combining several technical approaches to generate verifiable evidence that a question aligns with established educational taxonomies like Bloom's and SOLO.

The core of the framework is a structured certification workflow. It introduces a metadata schema to capture an item's provenance, predicted alignment with learning objectives, and any ethical flags. This data feeds into a practical 'traffic-light' system that automatically categorizes AI-generated questions: green for auto-certifiable, yellow for human review, and red for rejection. In a proof-of-concept study using 500 AI-generated computer science questions, the framework demonstrated feasibility, showing it could improve transparency, reduce manual instructor workload, and create an audit trail. The authors position this explainability and certification layer as an essential component for building AI assessment tools that institutions and accreditation bodies can actually trust and adopt at scale.