Exploring the use of generative artificial intelligence by university students: a systematic literature review

Full Text

Problem statement. Artificial intelligence (AI) has become a transformative force across various sectors, including education [1]. While AI has influenced learning environments and instructional processes over several decades, its impact remained gradual until the launch of OpenAI’s ChatGPT in 2022 [2; 3]. The release of ChatGPT marked a pivotal shift in the educational landscape, accompanied by rapid proliferation of other generative AI (Gen-AI) tools such as Perplexity.ai, BLOOM, ChatSonic, Claude, Bard, Whisper, and Jasper Chat. Collectively known as chatbots, these tools are built on large language models and process user input to provide interactive and contextually rich responses in both verbal and written forms [4]. Given their widespread adoption, Gen-AI tools have quickly become one of the most prevalent forms of AI in higher education, prompting a need to explore their implications on learning practices and outcomes. Gen-AI has gained popularity among key university stakeholders, such as faculty and students. Faculty members utilise these tools for developing instructional materials and giving feedback [5], facilitating personalised learning experiences [6; 7], and streamlining administrative tasks [8]. Students, on the other hand, leverage Gen-AI for diverse academic purposes, including research [9], writing assistance [10; 11], coding [12], and problem-solving [13]. Concurrently, the body of research on students’ use of Gen-AI significantly surpasses that on teachers’ applications. According to the Lens database, there are 2,305 papers focused on students, compared to 1,677 for faculty. Additionally, [2] report that 72 % of existing Gen-AI studies target student usage, while only 17 % concentrate on instructors. Despite this focus, existing literature reviews on AI intelligence demonstrate an underrepresentation of empirical studies focusing on university students’ use of Gen-AI [14-16]. This research focus highlights a need for a comprehensive examination of Gen-AI’s use. Addressing this gap is essential to developing a holistic understanding of Gen-AI’s role in higher education, particularly from the student perspective. Given the rapid evolution of Gen-AI technology along with its rapidly growing and often uncontrolled adoption among students, a systematic literature review is necessary to synthesise current knowledge. The present review aims to (1) map the current applications of Gen-AI by university students, (2) assess the impact on their learning processes and academic performance, and (3) identify research gaps and propose directions for future investigation. By focusing on high-quality publications from Q1 and Q2 Scopus-indexed journals, this review seeks to provide an in-depth panorama of Gen-AI’s integration into higher education. The literature identifies various uses of Gen-AI that can be classified into more complex dimensions. Thus, [10] begin the discussion by dividing Gen-AI tools into “mindless” functions, such as summarising and paraphrasing, and “mindful” functions, which can generate entire texts. The former is more popular among students. Building on this, [17] categorise Gen-AI applications based on the role of chatbots, viewing them as tools for language or idea generation or as writing partners. Also, [18] offer another classification based on how students interact with Gen-AI: one group makes no changes to AI-generated content, thereby learning little, while another group modifies the content according to their abilities. This idea is expanded by [11], who observe that “competent” writers use Gen-AI as a supplementary tool, whereas less competent writers rely more heavily on AI-generated text. [19] and [20] further explore the debate over whether to alter or directly copy Gen-AI feedback. [19] contrast the simple copy-paste approach with an iterative process of engaging with Gen-AI feedback, echoing [18] earlier findings. [19] also introduce two dimensions for classifying AI use: “content and component uses” and “structured adaptivity” versus “unstructured streamline”. [20] focuses on text-production patterns, classifying students into those who frequently modify Gen-AI feedback, those who explore Gen-AI resources, and those who prefer ready-made texts with citations. [21] also address the issue of editing AI-generated drafts, emphasising that interactivity is a part of a broader classification of Gen-AI applications. They divide Gen-AI usage into stages: “brainstorming and outlining”, “writing and revision”, and “feedback and evaluation”, aligning with the three main stages of the writing process: pre-writing, writing, and post-writing. Interestingly, [19] offer a similar classification based on writing stages and further differentiate between “hidden” and “predominant” Gen-AI uses, particularly in writing tasks. [22] identify similar Gen-AI uses and introduce a “metadimension” of seeking inspiration in interacting with AI, previously described as “overcoming writer’s block” by [17]. [23] expand these ideas by specifying Gen-AI uses at each writing process stage. Finally, while [24] discuss AI’s application in language practice, some patterns overlap with writing applications. In addition to speaking practice and personalised feedback, Gen-AI is used for grammar correction and finding appropriate words. While a substantial body of research has investigated the role of AI in administrative and instructional applications within higher education, studies specifically focusing on student engagement with Gen-AI tools are comparatively limited. Existing literature highlights Gen-AI’s capability to augment collaborative learning environments. For instance, [25] and [26] discuss Gen-AI potential to facilitate online collaborative debates and enhance writing processes through automated feedback mechanisms and question generation. [15] emphasise that AI-driven chatbots have proven effective in fostering group discussions and aiding students in articulating their perspectives more confidently. Further explorations in the literature reveal that AI tools not only support collaboration but also contribute significantly to skill development across various domains. For example, studies have documented the efficacy of chatbots in language learning contexts, where they enhance performance, critical thinking, empathy, communication skills, and overall student satisfaction [16; 15]. Moreover, Gen-AI applications extend to providing personalised academic assistance, where systems recommend resources and offer scaffolding tailored to individual learning needs [8; 26; 27]. Thus, previous literature reviews have broadly addressed the use of Gen-AI in higher education, yet detailed explorations of how university students utilise it in their studies still remain absent. To identify articles for inclusion in a literature review, various databases are utilised, which may vary depending on the research scope and objectives. For instance, [28] focused on one specific journal relevant to their topic. In contrast, [29] examined several leading journals on distance education, while [30] expanded their search to encompass journals in the broader field of higher education. In cases where the research scope is more extensive, all articles on selected topics from specific databases may be included. For example, [25] addressed three international databases: EBSCO Education Source, Web of Science, and Scopus. Similarly, [31] sourced articles from multiple electronic databases including Emerald, SpringerLink Journal, ScienceDirect Journal, SAGE, Taylor & Francis Online, and Wiley Online Library. [32] and [8] employed an even broader array of resources, including the Educational Database (ProQuest), Education Research Complete (EBSCOhost), ERIC (ProQuest), Scopus, Web of Science (Core Collection), and ProQuest Central. Similarly, the time frame for literature analysis is closely tied to the scope of the study. For example, when reviewing a single journal on a narrowly defined topic, as in the study by [28], the time frame may span several decades, from 1980 to 2014. In contrast, studies encompassing a few journals, such as [29], typically restrict the review period to a few years, in this case from 2014 to 2019. When a single topic is investigated across a few databases, the time frame is generally constrained to a decade or less; for instance, [33] analysed literature from 2007 to 2016 using databases like ERIC, Web of Science, Scopus, and PsycINFO. However, when multiple databases are employed to explore a single topic, as in [32], the time frame might be limited to just a few years. Generally, the reviewed studies adhere to the basic methodology and its variations as outlined by [34], which require a reviewer to establish a clear protocol for handling papers. This includes setting exclusion/inclusion criteria and following three primary stages of the review process: searching for papers, selecting them based on set criteria, and synthesising the findings. Concurrently, the tripartite approach advocated by [35] delineates the stages of a literature review such as description (a summary of the review content), synthesis (categorization of the reviewed research), and critique (evaluating the strengths and weaknesses of the literature). Furthermore, some studies utilise a critical literature review methodology aimed at identifying the most significant items in a field to produce a conceptual contribution [36]. This method includes a linguistically based approach where the underlying conceptualizations and assumptions are scrutinised, particularly the meanings of words [30]. The approaches to article analysis before synthesis vary across studies. For instance, [28] employ content analysis to examine the conceptual structure of text-based information and identify the most prevalent and recurring themes [37]. Some studies combine analytical methods like text mining and social network analysis [2], while others integrate content analysis [38] with thematic analysis [39; 32]. Another combination involves deductive analysis [40], where selected studies are categorised into pre-existing codes, and a constant comparative method used to inductively compare studies and create sub-themes that fit into larger themes on the topic [31]. The use of mixed methodologies is justified by the need to triangulate data, thereby enhancing the reliability and validity of research findings [41]. A popular method to report systematic review findings is preferred reporting items for systematic reviews and meta-analyses (PRISMA) guidelines, which include describing inclusion/exclusion criteria, defining a search strategy, screening and selecting articles, describing relevant studies, and analysing and synthesising the findings [31; 33]. While the PRISMA guidelines are widely adopted for systematic reviews, alternative methodologies such as DARE[7] and AMSTAR 2 [42] are also employed, providing diverse frameworks for conducting reviews with varying focuses and criteria, as exemplified by [16]. In systematic literature reviews, researchers employ a variety of tools to enhance their analysis and presentation of findings. For instance, Leximancer is utilised in [28] to generate concept maps from the titles and abstracts of selected papers, facilitating a visual exploration of key themes and concepts. Similarly, [29] apply Leximancer for text mining purposes, extracting and analysing textual data to identify patterns and relationships. For the analysis of social networks within the literature, tools such as Gephi and NodeXL are used by [29] to visualise and interpret the connections among authors, concepts, and publications. Furthermore, EPPI Reviewer 4.0 serves as a comprehensive platform for managing and analysing literature. This software is employed in studies like [25] and [16] to streamline the process of retrieving, reviewing, and coding papers, ensuring systematic data handling and enhanced accuracy in the synthesis of research findings. What exactly is analysed in found papers also varies. For instance, [28] focus solely on titles and abstracts to determine the relevance of articles to their study questions. [29] extend their analysis to reference lists, which can uncover additional relevant studies and extend the scope of their review. Meanwhile, in qualitative analyses such as the one conducted by [32], only those parts of articles that directly address the research questions are examined, allowing for a focused exploration of the topic at hand. Methodology. This study conducted a tertiary review utilising a systematic approach outlined in the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines, focusing on three key steps: search strategy and study selection, data analysis, and synthesis of findings. Data for this study was sourced from two databases: Google Scholar and Lens. The first one was employed to search for papers to analyse via qualitative analysis, the second - to perform a quantitative part. Google Scholar, a widely accessible and extensively used academic search engine, offers a broad spectrum of scholarly literature from various sources. Lens is a comprehensive scholarly database that aggregates global research across multiple disciplines, providing a robust platform for accessing a wide variety of scientific publications. These databases were chosen for their extensive coverage and accessibility, ensuring a comprehensive collection of relevant literature on AI use in higher education. Search string: “artificial intelligence” OR “machine intelligence” OR “intelligent support” OR “intelligent virtual reality” OR “chat bot*” OR “machine learning” OR “automated tutor” OR “personal tutor*” OR “intelligent agent*” OR “expert system” OR “neural network” OR “natural language processing” OR “smart technologies” OR “intelligent technologies” AND “higher education” OR “tertiary” OR “college*” OR “undergrad*” OR “graduate” OR “postgrad*” AND “learn*” OR student* AND “use/application” The following inclusion and exclusion criteria were applied to select papers from Google Scholar database. Inclusion criteria: • Source of publication: Articles indexed on Google Scholar. • Publication date: Studies published from January 2023 onward. • Language: Studies must be published in English. • Journal ranking: Only articles from journals ranked in the first or second quartile (Q1-Q2) according to relevant journal ranking metrics by Scopus. • Type of paper: empirical study. • Gen-AI is used by: students. Exclusion criteria: • Educational setting: Studies that do not focus on higher education contexts. • Relevance: Articles not addressing AI in higher education. • Data type: Excludes studies that do not present empirical data. • Peer review status: Excludes articles that are not peer-reviewed. The data was approached qualitatively: apriori and aposteriori codes were applied to the papers retrieved from Google Scholar. Mostly, the abstracts were analysed, but if the information was not sufficient, the results and discussion sections of the papers were addressed. Further to this, generic codes were used to extract basic factual information, such as the journal name and the authors’ countries. For a deeper analysis of the selected papers, we conducted a thematic analysis to identify recurring themes and patterns, as described by [39]. We employed a priori codes, derived from existing literature reviews on Gen-AI, to systematically categorise and analyse the data. Thematic apriori codes: - ways of using Gen-AI [8]; - impact of using Gen-AI [16; 26]; - Gen-AI as collaboration facilitator [15; 25]. This approach ensures that the coding process is grounded in established research, facilitating the comparison and synthesis of new findings with previous studies. Additionally, a posteriori codes that emerged during the analysis process were integrated to capture new themes and insights not previously identified in the literature. Thematic a posteriori codes: - Gen-AI adoption; - Gen-AI perception; - Gen-AI literacy; - preparation for Gen-AI use; - Gen-AI implementation; - Gen-AI performance in assignments; - skills required for collaboration with Gen-AI; - role of Gen-AI in collaboration; - alternative perspectives on AI and collaboration. Results and discussion. From the initial screening of 620 papers, 42 were selected for the final sample based on the predefined inclusion and exclusion criteria. The distribution of the selected papers across different years is as follows: 7 papers were published in 2022, 24 - in 2023, and 11 - in 2024. Among these, 6 were authored by a single individual, while 40 were collaboratively written with co-authors. In terms of publication quality, 27 papers were published in Q1 journals, and the remaining 15 in Q2 journals. A total of 27 journals were identified as sources of the included studies. The distribution of journals indicates a concentration of publications in specific outlets. The International Journal of Educational Technology in Higher Education emerged as the most frequently cited journal, with six articles, underscoring its pivotal role in disseminating research on AI application. This was followed by Computers and Education: Artificial Intelligence and Education and Information Technologies, each contributing significantly to the literature with four and three articles, respectively. Scientific Reports, Frontiers in Psychology, Cogent Education, Humanities and Social Sciences Communications and Education Sciences demonstrated repeated engagement in this research area with two articles published in each. The review identified a significant variance in the geographic distribution of authors contributing to the field. The United Arab Emirates (UAE) led with the highest number of authors, totalling 50. Following the UAE, the United Kingdom contributed 15 authors. China and the United States also showed considerable involvement, with 14 authors each. Other notable contributions came from Jordan and Australia, each with 8 authors. Indonesia, Lebanon, and Egypt contributed 6 authors each, demonstrating their ongoing commitment to the research area. The analysis also pointed out contributions from a range of other countries, with varying levels of involvement, including Saudi Arabia, Thailand, Poland, and Germany with 5 authors each. Several countries, such as Switzerland, Morocco, Israel, Malaysia, and Tunisia, each contributed one author. This diversity in authorship from countries across different continents emphasises the global relevance and interdisciplinary nature of the topic. To report findings based on a priori and a posteriori code that address the research questions posed in the methodology, a narrative summary of findings organised around key themes or factors will be employed, along with a narrative synthesis of the data as suggested by [43]. Research question 1. How do university students use Gen-AI in their studies? Gen-AI adoption (n=9)[8] Theories and models In measuring the adoption and acceptance of Gen-AI tools among university students, researchers commonly apply specific theories and models. The identified theories include Diffusion of Innovations Theory [44], Constructivism Learning Theory [45], Self-Determination Theory [46], Situated ExpectancyValue Theory (SEVT) [47], the Unified Theory of Acceptance and Use of Technology (UTAUT) [46], and UTAUT2 [48; 49]. The primary model identified is the Technology Acceptance Model (TAM), which forms the basis for UTAUT and UTAUT2 [46; 50]. Additionally, financial considerations [51] and academic major [52] were found to influence Gen-AI adoption. Factors and variables Student willingness to adopt Gen-AI is directly influenced by variables such as technological expertise and ease of doing business [44], as well as perceived ease of use and social influence [46]. Other significant factors include facilitating conditions [48], the benefits of using Gen-AI, supportive environments [47], user-friendliness [50], habit, performance expectancy, and hedonic motivation [49]. Key motivations for using ChatGPT are also its quick response time and ease of use, enhancing educational efficiency and task management [45]. Additionally, students with external funding are more likely to use Gen-AI than those who pay for it themselves [51]. Gen-AI is more readily accepted in engineering than in business, while the arts had lower adoption due to the tool’s text-centric nature [52]. Conversely, perceived usefulness, autonomy, and trust do not significantly affect Gen-AI acceptance [46], along with effort expectancy [48]. Although perceived usefulness did not directly impact adoption intention, it had an indirect effect through personalization (positive) and interactivity (negative) [50]. Ways of using Gen-AI (n=6) Gen-AI is widely applied in academic settings, facilitating a variety of tasks including coding [53; 54], administrative duties [53], and problem-solving [54]. Additional uses encompass clarifying subject concepts [54] and answering questions [53; 54]. Notably, the most common application of Gen-AI is in writing-related activities. Thus, during the pre-writing phase, Gen-AI assists in brainstorming [21], generating ideas [51; 53], conducting preliminary literature searches [54], outlining [21], and creating initial drafts [51]. It also supports literature studies [54] and data analysis [53]. In the while-writing phase, Gen-AI aids in the composition process itself [21; 54], enhancing language, style, and writing techniques [51], and performing translations [54]. Post-writing, Gen-AI is utilised for revisions, feedback, and evaluation [21]. The table below summarises the main uses of Gen-AI as identified in the analysed papers. Overview of Gen-AI applications identified in selected papers Gen-AI use Barrett, Pack, 2023 [21] Chan, Hu, 2023 [53] Dakakni, Safa, 2023 [51] Von Garrel, Mayer, 2023 [54] Brainstorming

About the authors

Anna E. Korchak

Higher School of Economics

Author for correspondence.
Email: aekorchak@hse.ru
ORCID iD: 0000-0002-6007-3098

Research Assistant, Centre of Sociology for Higher Education, Institute of Education

20 Myasnitskaya St, Moscow, 101000, Russian Federation

Yevgeny D. Patarakin

Higher School of Economics; Moscow City University

Email: patarakined@mgpu.ru
ORCID iD: 0000-0002-1216-5043
SPIN-code: 7044-4695

Doctor of Pedagogical Sciences, Professor at the Department of IT, Management and Technology, Institute of Digital Education, Moscow City University ; Professor at the Department of Educational Programmes, Institute of Education, Higher School of Economics

4/1 2nd Selskokhozyaystvenny Proezd, 129226, Moscow, Russian Federation; 20 Myasnitskaya St, Moscow, 101000, Russian Federation

Jamie Costley

United Arab Emirates University

Email: jcostley@uaeu.ac.ae
ORCID iD: 0000-0002-1685-3863

PhD, Assistant Professor at College of Education

H1 Sheik Khalifa Bin Zayed St, Al Ain, 15551, United Arab Emirates

References