Can coursework show how students used AI without relying on detection?
A two-course study made AI use, checking and revision more inspectable. Independent practical tests did not show a clear learning gain.
The short version
- What happened
- At one Chinese institution, 34 anatomy-and-physiology students received a traceable-assessment redesign and 33 students formed an intact comparison class; a second microscopy course included 15 students in six groups.
- Who it affects
- Students may be judged more fairly when assessment asks them to show what AI contributed, which claims they checked and why they revised their work, rather than treating an unreliable detector score as proof.
- What remains unproven
- One institution and one main professional course
Keep in mind
One institution and one main professional course
One intact class per condition with no random assignment
Teachers were involved in design and assessment
Feedback and revision support were bundled with traceability features
No clear difference on independent practical tests
No long-term retention or transfer beyond the courses
Source analysis · 1 linked source. Primary analysed source date: 26 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
The redesign collected evidence of use instead of guessing from the final answer
A peer-reviewed study replaced final-product-only assessment with a trail of AI-use statements, verification logs, attribution evidence, teacher checkpoints and explanations of revision. The main setting was an anatomy-and-physiology course at one Chinese biomedical-engineering programme: 34 students were in the redesign class and 33 in an intact comparison class. A smaller modern-microscopy adaptation involved 15 students working in six groups. There was no participant attrition.
Across three laboratory reports, the redesign class averaged 21.59 points compared with 17.91 out of 30, a mean difference of 3.68 points (95% CI 1.80–5.56; d = 0.96). Of that difference, 3.29 points were concentrated in disclosure and citation, verification and reflection—the dimensions deliberately changed by the redesign. AI-use statement uptake rose from 88.2% to 91.2% across implementation phases, while verification-log completeness rose from 1.44 to 1.74 on the study's scale.[1]
More visible work is not the same as better subject learning
The clearest independent check was the objective structured practical examination, or OSPE. Neither comparison was statistically significant: the first differed by 0.35 points (95% CI −0.10 to 0.80; p = 0.127) and the second by 0.19 points (95% CI −0.20 to 0.57; p = 0.334). An ethics-focused assessment also favoured the redesign class, but it was aligned with the redesign itself and cannot demonstrate broader ethical learning.
This was action research using one intact class per condition, not random assignment. Class membership was completely confounded with the condition, teachers helped design and assess the intervention, and checkpoints, feedback and revision opportunities changed together. Six exploratory traceability contrasts also lost statistical significance after correction for multiple testing. For teachers, the defensible conclusion is narrow but useful: process evidence can make AI-supported work easier to inspect. Confidence in learning effects needs multi-institution randomised or matched designs, blinded assessment, retention tests and transfer to new disciplinary tasks.[1]
The distinction that matters
What the assessments could—and could not—show
The strongest result concerned traceability, while the more independent practical outcome remained uncertain.
Course reports
Included AI-use and verification evidence
Higher design-aligned scores
Ethics task
Near-transfer assessment aligned to redesign
Not independent proof
OSPE practical tests
More independent subject-performance measure
No significant difference