How are universities redesigning assessment for GenAI?
A peer-reviewed review maps 72 studies to process, authentic and oral approaches. Most evidence is descriptive, so it does not yet show which redesigns improve learning or fairness.
The short version
- What happened
- The systematic review screened 486 records and included 72 peer-reviewed empirical studies from 24 countries and 20 disciplines.
- Who it affects
- Students may face more varied and meaningful assessment, but oral checks, process documentation and AI-supported grading can also add anxiety, workload and access barriers.
- What remains unproven
- English-only, title-restricted search; uneven regional and disciplinary coverage; mostly descriptive evidence; no meta-analysis or pooled learning effect.
Keep in mind
English-only, title-restricted search; uneven regional and disciplinary coverage; mostly descriptive evidence; no meta-analysis or pooled learning effect.
Source analysis · 1 linked source. Primary analysed source date: 16 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
The literature points to three redesign directions
A new systematic review finds universities responding to generative AI by making students' learning process more visible, using tasks closer to real practice and adding oral or dialogical elements. The authors screened 486 records and retained 72 peer-reviewed empirical studies published from January 2024 through June 2026. Those studies covered 24 countries and 20 disciplines. The synthesis offers educators a map of emerging practice, but it does not establish that any one design reliably improves learning, integrity or equity.
Process-oriented approaches ask for drafts, reflection, feedback uptake or other evidence of how work developed rather than grading only a polished final product. Authentic approaches use simulations, professional writing, workplace AI use and role-play. Oral and dialogical approaches include interviews, presentations, podcasts, screencasts, peer interaction and group projects. These are not entirely new ideas; the review's contribution is showing how GenAI has renewed and combined them across a fast-growing evidence base.[1]
The review also maps AI across the assessment cycle
Five educator uses emerged: creating assessments, giving students practice, grading, making learning visible and providing feedback. Creating tasks and feedback appeared most frequently. Examples ranged from generating exam questions to developing simulations, alternative formats and context-specific exercises. The practical attraction is clear: a general-purpose model can help a teacher prototype materials that once required specialist software or more preparation time. The risk is equally clear: generated tasks, grades and feedback can be inaccurate or poorly aligned with the intended learning outcome.
Two researchers independently reviewed full texts after duplicates were removed, reaching 98% initial agreement on eligibility and resolving disagreements by discussion. They then coded studies independently using deductive and inductive categories, reporting 97% agreement. Europe supplied 29 studies and Asia 25; Australia alone supplied 16 and the United Kingdom 15. Business, engineering, medicine and second-language education together accounted for 36 of 72 studies, so the apparent consensus reflects an uneven geographic and disciplinary base.[1]
The distinction that matters
Three ways assessment is being redesigned
The review groups observed approaches by the evidence of learning they try to make visible.
Process
Drafts, reflection and feedback
Shows how work developed
Authentic
Simulations and workplace tasks
Tests applied judgement
Dialogical
Oral, peer and group formats
Checks live understanding
A practice map is not an effectiveness ranking
The review included only English-language peer-reviewed journal articles whose titles matched the targeted GenAI, assessment and higher-education terms. That improves focus but can miss relevant work in other languages, grey literature or papers that use different terminology. More importantly, most included studies were qualitative, mixed-methods, descriptive or exploratory. The authors identified limited causal evidence and no pooled effect size. A theme occurring often is not proof that it works better than a conventional assessment or works equally well for every student.
For teachers, a sensible response is to choose the evidence of learning first, then decide where AI is allowed and how students will document its use. Oral checks can help verify understanding but may increase anxiety and create accessibility or scalability problems; process portfolios add evidence but also marking workload. The authors declared no competing interests, and the article reports no new dataset. Confidence would rise with randomised or strong quasi-experimental comparisons, longer follow-up, accessibility and workload measures, and studies from underrepresented regions. The review supports redesign beyond simple AI detection; it does not license universities to replace professional judgement with automated grading.[1]