Designing AI-Resilient Exams: Error-Correction Assignments in Engineering Education

Purpose

The article examines how redesigning portfolio exam tasks from generative writing to error-identification and correction can reduce the effectiveness of generative AI while preserving core learning outcomes and higher-order competencies in a STEM master’s course.

Design/methodology/approach

A field-experimental design in the master’s course “Applied Biorobotics” compares two traditional report-writing assignments with a redesigned assignment where students correct deliberately flawed reports. Four error-permuted versions limit collaboration. The grading focuses on error detection and correction. Results are complemented by questionnaires and heuristic LLM testing.

Findings

The new task format reduces the perceived usefulness of AI chatbots and can make AI use feel hindering, yet students still report achieving learning goals across Bloom’s revised taxonomy.

Research limitations

Based on one small STEM course and a single LLM benchmark, results are not fully generalizable but indicate how to redesign written exams to be more AI-resilient.

Practical implications

The study offers a concrete template for creating AI-resilient portfolio tasks while keeping grading feasible and authentic reasoning central.

Social implications

The approach models a middle path between banning AI and unrestricted use, supporting responsible AI engagement and helping preserve trust in university assessment.

Originality/value

The article empirically tests an operationalized AI-resilient portfolio design, showing that open, asynchronous exams can still foster higher-order learning in AI-rich environments.