What do 394 higher-education studies show about generative AI?
A peer-reviewed scoping review finds that universities mostly used generative AI for feedback, tutoring and writing. The evidence is substantial but still dominated by surveys and short-term measures, with little attention to equity or learning processes.
The short version
- What happened
- The authors searched EBSCOhost and Web of Science in January 2026, screened 2,264 unique records and included 394 empirical studies of enacted, task-anchored generative-AI use.
- Who it affects
- Students are already encountering AI mainly through feedback, tutoring and writing, but current research says less about whether they retain knowledge or can perform independently later.
- What remains unproven
- The scoping review mapped studies but did not assess their quality or pool effect sizes.
Keep in mind
The scoping review mapped studies but did not assess their quality or pool effect sizes.
Only English-language, peer-reviewed journal articles indexed in EBSCOhost or Web of Science through December 2025 were eligible.
Most included research focused on students; only 15 studies focused on instructors and 26 on both groups.
Surveys and short-term measures were common, while logs, observation and expert ratings were comparatively rare.
Study-level categories cannot show whether every reported use or outcome was beneficial, harmful or null.
Source analysis · 1 linked source. Primary analysed source date: 15 October 2026.
Read the full report · 10 min The evidence, methods, implications and limitations
The answer: widespread classroom use, but not a verdict that AI improves learning
The strongest finding is about what universities have studied, not whether generative AI raises attainment. Across 394 empirical studies, the most common enacted uses were feedback or tutoring and writing support. Researchers most often targeted engagement, usability, critical thinking, subject knowledge, academic performance, writing quality and AI literacy. Self-regulated learning, ethics, creativity, feedback quality and equity were less visible.
That distinction matters for teachers and university leaders. A large literature can still be unable to answer a simple causal question if it combines different tools, courses, outcomes and study designs. This review is a map of that terrain. It shows where experimentation and measurement have concentrated, but it does not calculate an average learning effect or show that a particular chatbot, course design or policy works across institutions.
The paper therefore cuts through two opposing headlines. The evidence base is no longer merely speculative: hundreds of studies examined actual course-related use. But the existence of 394 studies is not 394 replications of one intervention. The safer conclusion is that higher education has accumulated substantial experience with bounded tasks while evidence about durable learning, unequal effects and the learning process remains thinner.[1]
How the reviewers reached 394 studies
The authors followed the PRISMA extension for scoping reviews. In January 2026 they searched EBSCOhost and Web of Science Core Collection for English-language, peer-reviewed journal articles published from November 2022 through December 2025. The search returned 2,778 records. After 514 duplicates were removed, 2,264 unique records went through title and abstract screening, and 523 articles reached full-text review.
Inclusion required original empirical research in higher education with generative AI embedded in an authentic teaching or learning activity: an assignment, assessment, feedback cycle, instructional intervention or instructor-led class activity. The team excluded output-only benchmark studies, opinion and review articles, work outside higher education, and perception surveys where no enacted course use was documented. Full-text assessment produced 394 included studies.
Three researchers reviewed the full texts and resolved borderline cases by consensus. They jointly calibrated coding on ten included studies, then recorded study design, participant focus, tool, use case, targeted outcomes, assessment methods, stakeholder perceptions and governance indicators. This is a transparent and appropriately broad method for describing a fast-moving field, although it does not provide the risk-of-bias assessment expected from a review trying to estimate effectiveness.[1]
Most research examined students and familiar academic tasks
Mixed-methods research was the largest design group, with 213 studies or 54% of the corpus. Another 109 studies were quantitative and 72 qualitative. Participant focus was heavily tilted toward students: 352 studies focused on students, 15 on instructors, 26 on both, and one on several stakeholder groups. That imbalance limits how confidently the literature can describe staff workload, assessment validity or institutional implementation.
Feedback and tutoring appeared most often across tool categories. Writing support was the other major activity, followed by language learning or translation, lesson planning, assessment or grading, general academic help, programming or debugging, content creation, simulation and reflection. The pattern suggests that universities have mainly fitted generative AI into existing tasks rather than reorganised curriculum around new human-machine roles.
The paper's cross-tabulations make that concentration concrete. Feedback or tutoring was linked to 51 studies targeting engagement, 46 targeting achievement, 45 targeting usability, 45 targeting subject knowledge and 37 targeting critical thinking. Writing support was linked most often to writing quality, in 50 studies. Those are counts of coded studies, not pooled effect estimates; a study appearing in a category may report benefit, harm, mixed findings or no meaningful difference.[1]
Measurement leant heavily on surveys and end products
Across the 394 studies, the reviewers recorded 891 assessment-method codes because one study could use several methods. Survey scales appeared in 269 studies, or 68.3%. Tests, quizzes or rubrics appeared in 185 studies, artifact analysis in 167 and interviews or focus groups in 143. Analytics or logs appeared in 76 studies, observation in 30 and expert ratings in only 13.
Surveys are legitimate tools for measuring experience, motivation and perceived usefulness. Rubrics and artifact analysis can show whether a submitted product meets defined criteria. Neither automatically reveals how a learner reasoned, checked an answer, revised a prompt or transferred knowledge to an unfamiliar task. A polished essay or higher satisfaction rating can coexist with weaker independent understanding.
For practitioners, the gap is practical rather than academic. If a university wants students to develop critical thinking or self-regulation, it should measure the process: attempts, revisions, source checking, error correction and performance after AI access ends. The review found those process-sensitive approaches were comparatively uncommon, leaving prompting, feedback uptake, verification and over-reliance only partly visible.[1]
Students and instructors were mostly ambivalent, for different reasons
Among 315 student-related perception cases, 228 or 72.4% were coded as mixed, 85 as positive and two as negative. Among 35 instructor-related cases, 28 or 80% were mixed and seven positive. These categories summarise how studies reported views; they are not a representative opinion poll of all students or staff.
Student themes most often concerned learning and skill development, engagement and reliability or usability. Instructor themes centred more on pedagogy and assessment, efficiency and workload, professional learning, reliability, ethics and integrity. A student may value faster drafting or feedback while an instructor worries that the same shortcut makes it harder to judge independent understanding. Both positions can be rational within the same course.
Governance also looked less developed than general concern. The review coded 151 studies as recommending institutional policy and 148 as mentioning ethics or policy, but only 52 as reporting instructor-level rules in the enacted learning context; 43 mentioned none. A warning in a policy document is not the same as a course rule that tells students what assistance is permitted, how to disclose it and which work must remain unaided.[1]
Equity received too little direct attention
Equity, inclusion and access were among the less frequently targeted outcomes. Equity themes appeared in 18 student-perception cases and three instructor cases, while privacy and data security appeared in 18 student cases and two instructor cases. Low frequency does not show that inequity is absent; it shows that the reviewed research rarely made it a central measured outcome.
That blind spot matters because access to paid models, device quality, language support, disability accommodations, discipline-specific norms and AI literacy can change who benefits. A tool that improves the average product may still widen a gap if some students receive better models, know how to verify outputs or can afford repeated use. Conversely, well-designed language or accessibility support may help students whom conventional instruction underserves.
Universities testing generative AI should therefore report results by relevant student groups where sample size and privacy allow, document which model tier was used and provide an equivalent route for students who cannot or do not wish to use it. Equity should be a planned outcome, not a paragraph added after a pilot is complete.[1]
What teachers and university leaders can act on now
Treat generative AI as an instructional-design decision, not a software switch. Start with the learning goal, specify the role AI may play, preserve a route to human feedback and decide what evidence will demonstrate independent understanding. In writing, that might mean retaining drafts, source checks and a short oral explanation. In programming, it might mean debugging logs plus an unaided transfer task.
Teachers also need explicit, local rules. Students should know whether idea generation, editing, code completion or full drafting is allowed; how use must be disclosed; what data must not be entered; and when an output must be checked against a primary source. Institutions should support that work with staff development and discipline-sensitive guidance rather than one generic statement for every course.
Evaluation should extend beyond satisfaction and immediate grades. Useful measures include retention after AI access ends, transfer to new problems, the quality of revisions, error detection, time spent by staff correcting outputs, and outcomes by prior attainment or access. The review was supported by two Social Sciences and Humanities Research Council of Canada grants, and the authors declared no competing interests.[1]
Evidence limits and what would change the assessment
The review searched two large databases but limited inclusion to English-language, peer-reviewed journal articles through December 2025. It may miss studies in other languages, conference proceedings, newer 2026 work and evidence not indexed in those databases. Coding compressed complex courses into categories, and multi-label assessment methods were counted at study level rather than matched one-to-one with every claimed outcome.
Most importantly, the authors did not assess study quality or calculate effect sizes. This article cannot tell a dean that generative AI raises marks by a particular amount, or that mixed perceptions imply an acceptable risk-benefit balance. Its value is diagnostic: it shows a literature concentrated on familiar tasks and relatively easy-to-collect measures while process, governance and equity evidence lag.
Confidence that generative AI improves meaningful learning would rise with preregistered, multi-institution studies using common comparators, independent assessments, longer follow-up and transparent model versions. Stronger evidence would connect particular course designs to retention, transfer, staff workload and unequal effects. Until then, the review supports careful, measurable experimentation—not a blanket mandate to adopt or ban generative AI.[1]