Generative AI Is Everywhere on Campus, But Students Say Support Is Lagging

Generative artificial intelligence has become so deeply woven into student life that nearly eight in ten students now use it several times a week or every single day. That is one of the striking findings of a new exploratory survey published in Discover Education, which examined how students in Indonesian, English, and Hungarian linguistic contexts use AI tools and how supported they feel by their institutions. The study, led by Deddy Gusman of the University of Debrecen and Universitas Pahlawan Tuanku Tambusai together with colleagues, surveyed 186 students and found that 78.0 percent reported using generative AI several times per week or daily, with a 95 percent confidence interval running from 71.5 to 83.3 percent. Even more remarkable, 96.8 percent of respondents reported using ChatGPT specifically, while 79.6 percent said the technology had improved their learning.

The scale of adoption documented in the study leaves little doubt that generative AI is no longer an experimental novelty in higher education. Of the 186 respondents, 185 reported using at least one of the four major general-purpose large language models: ChatGPT, Gemini/Bard, Claude, or DeepSeek. ChatGPT was by far the dominant tool, named by 180 of the 186 participants, followed by Gemini/Bard at 53.8 percent, Grammarly at 31.7 percent, DeepL at 31.2 percent, and Claude at 18.3 percent. The researchers deliberately included adjacent tools such as Grammarly and DeepL in their checklist, but their analysis showed that excluding them would barely change the picture, since almost everyone who used a writing or translation aid also reported using a general-purpose language model.

What students do with these tools is equally revealing. The most frequently reported academic purposes were research or information seeking at 86.0 percent, problem solving at 71.0 percent, understanding concepts at 65.6 percent, writing assignments at 63.4 percent, and language translation at 54.8 percent. Because respondents could select multiple purposes, these figures describe overlapping repertoires rather than mutually exclusive categories. The most common purpose pairs were research combined with problem solving, chosen by 65.6 percent of respondents, and research with concept understanding, chosen by 60.8 percent. The data suggest that students treat AI less as a shortcut for producing finished work and more as a versatile companion for the entire arc of academic labor, from the first search for information to the final polish of a draft.

The study’s conceptual centerpiece is what the authors call practice–governance decoupling: the coexistence of routine, everyday AI use with uneven and often unclear institutional support. Students clearly perceive the gap. When asked what forms of institutional assistance they most needed, 37.1 percent chose free institution-paid access to AI tools and 24.2 percent chose clearer institutional AI policies, making these the two most frequently requested forms of support. When asked about the biggest barrier to AI adoption, 31.2 percent selected educational policies or training, ahead of ethical concerns at 19.4 percent, economic barriers at 18.3 percent, cultural barriers at 17.2 percent, and technical barriers at 14.0 percent. In other words, the obstacle students identify most often is not the technology itself but the absence of guidance about how to use it well.

Attitudes toward AI were far from uncritically enthusiastic. While 79.6 percent of respondents perceived that AI had improved their learning, views on the overall balance of the technology were mixed: 51.6 percent described it as equally beneficial and risky, 23.7 percent as more beneficial than risky, and 12.9 percent as essential for future success, while the two predominantly negative categories together accounted for 11.8 percent. This ambivalence mirrors the wider literature on student perspectives, which consistently finds that students value AI for speed, feedback, and confidence while remaining worried about misinformation, unfair advantage, overreliance, privacy, and academic integrity. The respondents themselves rated their agreement that using AI constitutes cheating at a low mean of 2.17 on a five-point scale, suggesting that most do not see routine use as inherently dishonest.

One of the most intriguing patterns in the data involves the people around the student. Classmate support received the highest mean rating of the seven support-related items, at 3.94 on a five-point scale, followed by self-rated effective use at 3.79 and ethical use at 3.75. Professor encouragement, by contrast, averaged just 3.14, a full four-fifths of a point below peer support. In the exploratory comparison between Indonesian and Hungarian residents, this professor-encouragement gap was the largest contrast in the entire study, with Indonesian residents reporting a mean of 3.41 against 2.18 for Hungarian residents, yielding a signed rank-biserial effect estimate of 0.549. By comparison, the contrast in perceived learning improvement between the two groups was tiny, at just 0.050. The authors are careful to stress that these are perceptions, not measurements of actual teaching behavior, and that the descriptive difference does not establish any temporal sequence of peer versus faculty adoption.

The methodological honesty of the paper is notable and worth understanding, because it shapes what the findings can and cannot tell us. The questionnaire existed in three language versions sharing a common set of core items, translated from English into Indonesian and Hungarian by native speakers with independent bilingual back-translation. Yet the authors explicitly state that measurement invariance was not established: translation procedures sought semantic correspondence but did not guarantee psychometric equivalence. Recruitment was non-probability and self-selecting, distributed through WhatsApp groups for the Indonesian and English versions and in classes for the Hungarian version. Country of residence, survey language, recruitment route, and response timing all overlapped heavily. All 21 Hungarian-language responses arrived on a single day, 20 May 2025, while Indonesian-language responses trickled in from late May to early November. The study was not preregistered and no multivariable model was fitted.

These constraints matter most when interpreting the Indonesia–Hungary contrasts. A sensitivity analysis comparing the small subset of respondents who chose the English-language questionnaire, 11 Indonesian residents and 14 Hungarian residents, found that the large professor-encouragement contrast shrank considerably and the coded policy-item contrast nearly vanished. The authors interpret this as evidence that the full-sample contrasts depend strongly on who was sampled, without being able to divide the differences into language, country, recruitment, or timing components. They also caution that the policy-climate question mixed degrees of restrictiveness with an unclear/no-policy option, making its numerical coding unreliable as a measure of policy clarity. Fifty-six point seven percent of Indonesian residents described their institutional policy climate as somewhat or very permissive, compared with 17.9 percent of Hungarian residents, while 38.5 percent of Hungarian residents selected the unclear/no-policy option.

The study also speaks to a growing concern in the literature about a so-called GenAI divide in higher education. Recent work shows that readiness, access, perceived advantages, and policy support vary across demographic groups, institutional settings, and cultural contexts, raising the possibility that the benefits of generative AI are unevenly distributed. Within the present sample, some subgroup patterns were suggestive: undergraduates reported the highest classmate-support and effective-use means, while nearly half of the doctoral respondents, 15 of 31, selected the unclear/no-policy option for their institutional climate. Educational policy or training was the most frequently selected barrier for both women, at 32.6 percent, and men, at 30.0 percent. These subgroup results are purely descriptive and hypothesis-generating, given the small group sizes.

The authors ground their interpretation in several complementary theoretical traditions: technology-acceptance research with its concepts of perceived usefulness and facilitating conditions, the domestication tradition examining how technologies enter everyday life, diffusion theory, and institutional theory, including the distinction between policy–practice and means–ends decoupling drawn by Bromley and Powell. From these they build a four-component heuristic linking student practice, perceived institutional conditions, the practice–governance configuration, and proposed outcomes for future evaluation, while emphasizing that no causal paths were estimated. The practical takeaway is immediate and concrete: students are telling institutions, clearly and consistently, that what they want most is funded access to approved AI tools and clearer policies on acceptable use. Whether particular institutional responses, from faculty development to assessment redesign, actually improve educational outcomes remains an open question that will require longitudinal and intervention studies. What this survey establishes is the starting point: routine AI use has arrived in students’ academic lives, and the support structures around it have not yet caught up.

Subject of Research: Student use of generative AI and perceived institutional support in higher education across Indonesian, English, and Hungarian linguistic contexts

Article Title: Student use of generative AI and perceived institutional support across three linguistic contexts

Article References: Gusman, D., Aditya, D. S., Putri, D. W., Damayanti, N., Nurmalina, N., & Bujdosó, G. (2026). Student use of generative AI and perceived institutional support across three linguistic contexts. Discover Education, 5(1), Article 1175. https://doi.org/10.1007/s44217-026-02325-9

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02325-9

Keywords: generative AI, higher education, ChatGPT, large language models, AI governance, student perceptions, trilingual survey, institutional policy, practice-governance decoupling, educational technology, measurement invariance, AI in education