Beyond AI detection: traceable assessment design for generative AI-supported professional coursework

Generative artificial intelligence (GenAI) challenges final-product-only coursework assessment: a polished report or design proposal may no longer show how students used AI, which claims they checked, how they attributed sources and AI contributions, or what they revised. This study examined traceable assessment design as a practical response. Rather than relying on AI detection or broad permission rules, assessment tasks were redesigned to collect evidence of students’ AI use, verification, attribution, feedback-mediated revision, and reflective judgement. Using a multi-cycle action research design with an embedded mixed-methods approach, we studied two professional courses: anatomy and physiology as the main evaluation setting and modern microscopy as a secondary adaptation setting. Data sources included coursework, AI-use statements, verification logs, assessment records, student interviews, teacher redesign logs, and traceability-related process indicators. In the main setting, the redesign class recorded higher cumulative report scores than the comparison class (21.59 vs 17.91/30; mean difference = 3.68, 95% CI [1.80, 5.56], Welch t (63.1) = 3.91, p < 0.001, d = 0.96). Of this difference, 3.29 points (89.4%) were concentrated in the redesign-targeted disclosure/citation, verification, and reflection dimensions, compared with 0.39 points across physiology accuracy and the data/evidence chain. The redesign class also recorded higher scores on a design-aligned ethics-focused assessment (14.53 vs 12.36/20; mean difference = 2.17, 95% CI [1.08, 3.26], Welch t (63.9) = 3.99, p < 0.001, d = 0.98), which was not independent evidence of broader ethical learning. Neither OSPE comparison was statistically significant ( p = 127 and 0.334). Across implementation phases, AI-use statement uptake was 88.2% and 91.2%, and verification-log completeness was 1.44 and 1.74. In microscopy, group report scores across three technically different assignments (7.60, 8.03, and 8.60/10) were interpreted descriptively, not as longitudinal improvement. Traceable assessment thus made students’ documented GenAI use, verification, attribution, and revision more inspectable without establishing improved disciplinary learning.