Higher education assessments in the age of GenAI: a systematic review

Synthesis: Crompton and Burke's (2026) PRISMA review of 72 empirical studies published between 2024 and 2026 maps how generative AI is being absorbed into higher education assessment rather than how it is being rethought. The evidence base is wide — 24 countries and 20 disciplines — but thin in method: qualitative designs dominate and most studies examine a single course. Three design responses recur: process orientation, authenticity and oral and dialogical approaches. Five uses of GenAI across the assessment cycle are documented — creating assessments, practice, grading, transparency of learning and feedback — and the review's central argument is that these are being layered onto existing assessment principles of validity, authenticity and constructive alignment rather than displacing them.

Key Findings

1. The review's scope and size. 72 studies met the inclusion criteria for empirical research published between 2024 and 2026, analyzed with PRISMA. The authors position this as an update specifically on assessment, noting that prior syntheses examined GenAI in teaching and learning more broadly and left the assessment-specific picture less clear.
2. Two research purposes dominate the field. Studies either investigate the use of GenAI in assessment or examine perceptions of it. Within use-focused work the authors identify two axial codes — evaluation of new assessment approaches and assessment redesign — while perception-focused work splits into student and faculty perceptions.
3. Qualitative approaches are the most common methodology. The review reports that qualitative designs were the most frequent research methodology across the corpus, which the authors read as a sign of an exploratory field rather than one testing interventions at scale.
4. Three assessment design approaches emerged in response to GenAI. Process orientation shifts evidence toward how work develops; authenticity reframes tasks toward meaningful disciplinary problems; oral and dialogical approaches move assessment toward live demonstration and dialogue. These are described as responses to concerns about authorship and the evidentiary value of submitted products, not as a rejection of written assessment.
5. Five uses of GenAI span the assessment process. Educators used GenAI to create assessments, to provide practice for assessments, to grade, to make learning more transparent, and to generate feedback. The review presents these as the operative affordances that studies actually investigated, rather than as recommendations.
6. Geographic and disciplinary concentration is uneven. The 72 studies span 24 countries and five continents; Australia and the United Kingdom together account for 31 of them. Business, engineering, medicine and second-language education account for roughly 47% of the disciplinary spread.
7. Assessment scholarship is the frame, not GenAI. The authors argue GenAI's influence should be read within the existing trajectory of assessment theory — validity, authenticity and alignment with learning outcomes — rather than treated as a break from it, and they direct their synthesis at researchers, educators, institutional leaders and policymakers.

What the three design responses actually change

The review treats process orientation, authenticity and oral or dialogical approaches as distinct but converging moves away from the single submitted product. Process orientation is the broadest: it reallocates evidence toward drafts, revision histories and staged work, which means the assessment question becomes when evidence is collected rather than only what is collected. Authenticity is the most contested, because "authentic" is used in the studies to mean both professional-task realism and disciplinary judgment, and the review does not resolve which sense the field intends. Oral and dialogical approaches are the most direct response to authorship uncertainty: they relocate part of the evidence into live performance or conversation, where the reasoning behind a claim becomes observable. The review's own framing connects these to academic integrity concerns without claiming that integrity is the only driver — several studies describe redesign as pedagogical rather than defensive.

The five uses and where the evidence is weakest

Creating assessments, practice, grading, transparency and feedback are documented unevenly. Feedback and practice are the best evidenced, which aligns with the broader GenAI-in-education literature: these are the uses closest to existing instructional practice and easiest to study in a single course. Grading and transparency are the most consequential and the least examined, because both raise questions about validity and about what students are entitled to know concerning how AI contributed to a judgment. The review notes that the studies rarely report the AI Governance or disclosure arrangements under which AI-assisted assessment operated, which makes it difficult to compare designs across institutions. Transparency of learning is also the use most exposed to disclosure norms, since it sits at the boundary between making criteria visible to students and making AI's role visible in the assessment process.

Reading the findings against assessment theory

The review's strongest contribution is methodological rather than empirical. By organizing the corpus around long-standing assessment principles — validity, authenticity, alignment — it makes the field's implicit assumptions visible: most studies assume that using GenAI well means preserving the constructs that assessment already valued. That assumption is defensible but untested. It also means the review inherits the limits of its sources: short-term, single-course, self-report-heavy designs that the authors themselves flag, alongside a geographic and disciplinary spread that leaves large parts of global higher education unrepresented. The authors are explicit that the field's next step is not more descriptions of use but evidence about whether redesigned assessment produces better learning, and for whom.

What this means for practice

• Instructors. Treat the three design approaches as a menu of levers rather than a mandate. If you are already collecting process evidence, authenticity and oral components can be added incrementally; if you are not, adding all three at once changes the assessment's construct, not just its security.
• Assessment designers. Ask which of the five uses your design depends on and whether the evidence for that use actually exists in your context. Feedback and practice have the most supporting research; grading and transparency have the least, so those are the places to pilot rather than assume.
• Faculty developers. Use the review's two-purpose framing to help instructors locate their own question — are they studying use or perceptions? — because the two call for different evidence and different claims about impact.
• Institutional leaders and policymakers. The reported absence of governance and disclosure detail in the studies means institutional policy cannot be derived from this literature yet. Treat the review as a map of what has been tried, not as a validated playbook.

Limitations

• The review's own evidence base is short-term and single-course. The authors state that most included studies examined one course or one cohort, so findings describe designed interventions rather than institutional practice.
• Qualitative dominance limits causal claims. With qualitative designs the most common methodology, the review can document how GenAI was used in assessment but cannot establish that its use improved learning outcomes.
• Geographic and disciplinary coverage is concentrated. Australia and the UK supply 31 of 72 studies, and four disciplines supply roughly 47% of the disciplinary spread, so generalization beyond those settings is not supported.
• GenAI capability moved during the review window. Studies published in 2024 examined earlier model generations than those published in 2026, and the review does not separate findings by tool vintage.
• Governance and disclosure practices are under-reported. Because studies rarely describe the policy context in which AI-assisted assessment occurred, comparisons across institutions and systems are difficult to make.

Citation

Crompton, H., & Burke, D. (2026). Higher education assessments in the age of GenAI: a systematic review. International Journal of Educational Technology in Higher Education, 23(1).