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

A polished report or design proposal no longer tells educators what it once did. With generative artificial intelligence now able to produce fluent, well-structured text on demand, a final submission may conceal how a student actually used AI tools, which claims they checked against sources, how they attributed borrowed material, and what they revised after feedback. A study published in Humanities and Social Sciences Communications argues that the practical response is not better detection software or broader permission rules, but a redesign of assessment itself so that the process of AI-supported work becomes visible and inspectable.

The researchers call this approach traceable assessment design. Instead of treating the finished product as the sole object of grading, the redesigned tasks collected documented evidence at each stage: statements declaring how generative AI was used, logs recording which claims were verified and against what sources, explicit attribution of both human and AI contributions, revisions made in response to feedback, and reflective commentary on the student's own judgement. The aim was to make AI use, verification, attribution and revision something an assessor can actually examine, rather than something to be policed after the fact.

The study used a multi-cycle action research design with an embedded mixed-methods approach, drawing on coursework, AI-use statements, verification logs, assessment records, student interviews, teacher redesign logs and traceability-related process indicators. Two professional courses served as settings. Anatomy and physiology was the main evaluation site, where a redesign class was compared with a comparison class; modern microscopy acted as a secondary setting to test whether the approach could be adapted to a technically different subject.

In the anatomy and physiology setting, the redesign class recorded higher cumulative report scores than the comparison class: 21.59 versus 17.91 out of 30, a mean difference of 3.68 points with a 95 percent confidence interval of 1.80 to 5.56, Welch t(63.1) = 3.91, p < 0.001, and an effect size of d = 0.96. The composition of that difference is the study's most telling result. Of the 3.68-point gap, 3.29 points, or 89.4 percent, were concentrated in the dimensions the redesign directly targeted: disclosure and citation of AI use, verification, and reflection.

Only 0.39 points fell across physiology accuracy and the data and evidence chain. In other words, the redesign class did not show a broad gain in disciplinary content knowledge. The advantage lay almost entirely in the practices the new assessment made explicit and rewarded: documenting AI use, checking claims, and reflecting on the work.

The authors are careful about this distinction, noting that traceable assessment made students' documented GenAI use, verification, attribution and revision more inspectable without establishing improved disciplinary learning. The redesign class also scored higher on a design-aligned, ethics-focused assessment, 14.53 versus 12.36 out of 20, a mean difference of 2.17 points with a 95 percent confidence interval of 1.08 to 3.26, Welch t(63.9) = 3.99, p < 0.001, and d = 0.98. The researchers again qualify the interpretation: because this assessment was aligned with the redesign, the result is not independent evidence of broader ethical learning.

Two comparisons on the Objective Structured Practical Examination, a practical skills assessment, were not statistically significant, with p values of 0.127 and 0.334, meaning practical examination performance did not differ measurably between the groups. Process indicators showed that the redesigned tasks were actually used. Across implementation phases, uptake of the AI-use statement reached 88.2 percent and then 91.2 percent of students, and verification-log completeness rose from 1.44 to 1.74 on the study's measure.

These figures indicate that most students completed the disclosure statements and that the logs capturing verification activity became more complete as the design matured through successive cycles. In the modern microscopy course, the approach was adapted to three technically different assignments, with group report scores of 7.60, 8.03 and 8.60 out of 10. The researchers interpret these numbers descriptively only, explicitly not as longitudinal improvement, since the assignments differed technically and no comparison group was involved.

The secondary setting demonstrates transferability of the design rather than an outcome effect. The broader implication is a shift in what assessment can credibly claim to measure. Traceable assessment reframes the evidence base: disclosure statements, verification logs, attribution practices, feedback-mediated revisions and reflective judgement become the graded record.

The study's own caveats matter as much as its statistics. The score advantages were concentrated in redesign-targeted dimensions, ethics gains were not independently established, practical examination results were flat, and the microscopy findings were descriptive. What the evidence supports is narrower but concrete: process-based redesign can make AI-supported professional coursework more transparent and inspectable, offering educators a documented alternative to the arms race of AI detection.

Subject of Research: Educational assessment and generative AI

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

Article References: Li, X. (2026). Beyond AI detection: traceable assessment design for generative AI-supported professional coursework. Humanities and Social Sciences Communications. https://doi.org/10.1057/s41599-026-09634-2

Image Credits: AI Generated

DOI: 10.1057/s41599-026-09634-2

Keywords: generative AI, assessment design, higher education, academic integrity, action research, traceable assessment, Beyond, detection, traceable, assessment, design, generative