Artificial Intelligence in Higher Education: A Comparative Bibliometric Study of Korean and Chinese Research (2016-2025)

2.1 AI and Generative AI in Higher Education

Before generative AI became widely available as an academic tool, AI research in higher education was largely organized around application-oriented and system-level concerns. Earlier studies identified major areas such as profiling and prediction, assessment and evaluation, adaptive systems, personalization, intelligent tutoring, AI assistance, and student learning management [1, 2]. Those studies provide a baseline for the present research by clarifying the dominant concerns of higher education AI research before the widespread diffusion of generative AI.

The emergence of generative AI has expanded the scope of higher education AI research. Earlier AI applications were often embedded in institutional systems or specialized educational technologies. Generative AI, however, can be used directly by students and teachers for writing, feedback, information search, research support, teaching preparation, and assessment-related tasks. Recent literature emphasizes that generative AI in education raises questions not only about technological adoption, but also about curriculum design, teaching and learning practice, assessment, academic integrity, institutional guidelines, and policy responses [3, 4]. Higher education AI research has consequently moved beyond system-level support toward academic practice, knowledge production, responsible use, and educational governance.

Policy and educational contexts further shape the development of higher education AI research. Korea and China have both engaged with digital education and technology-related higher education reform, but their efforts have been organized through different policy priorities and institutional arrangements. Earlier comparative work on ICT policies in Chinese and Korean higher education suggests that the two systems provide a useful basis for examining how technology-related educational reforms are shaped by different policy environments [21]. In Korea, higher education has been incorporated into broader national strategies for digital and artificial intelligence talent development. The 2022 Comprehensive Plan for Digital Talent Development included university and graduate-level pathways as well as digital convergence education for students from diverse disciplinary backgrounds [9]. At the university level, the Digital New Technology Talent Development Innovation Sharing University initiative supported university consortia in eight advanced fields, including artificial intelligence, to jointly develop and share curricula [10]. In China, policy documents such as the New Generation Artificial Intelligence Development Plan and the Artificial Intelligence Innovation Action Plan for Institutions of Higher Education link AI more directly to national innovation strategy, higher education reform, talent cultivation, education digitalization, and institutional transformation [16, 17]. These contexts support the need to examine Korean and Chinese higher education AI research as contextually situated bodies of scholarship rather than as directly comparable publication counts.

Across higher education systems, similar technological developments may be organized through different scholarly problem structures. Current higher education AI research covers teaching and learning, academic practice, digital transformation, assessment, responsible use, institutional response, and emerging forms of AI-related literacy. The present study therefore examines how these concerns are organized within Korean and Chinese journal literatures, rather than assuming in advance which themes should be central in each country.

2.2 Bibliometric Mapping of National Journal Corpora

Bibliometric mapping is appropriate for examining fast-developing research fields because it can identify publication trends, keyword structures, co-occurrence networks, thematic clusters, and emerging topics across a large body of literature. Donthu et al. noted that bibliometric analysis helps researchers map the intellectual structure and evolutionary patterns of a field [22]. CiteSpace was designed to detect and visualize emerging trends and transient patterns in scientific literature [23]. In fields such as AI and generative AI in education, bibliometric methods can describe publication growth and show how research topics are connected, reorganized, and transformed over time.

Previous bibliometric and review studies have begun to map AI and generative AI in education. Liu et al. used CiteSpace and VOSviewer to analyze Web of Science records. They reported rapid growth in research on generative AI in education, identified major journals, collaboration patterns, and research topics, and noted concerns about academic integrity and educational inequality [19]. Dai et al. reviewed 2,762 Web of Science articles published between 2022 and 2025. They reported rapid growth in the field, led by the United States, China, and the United Kingdom, alongside increasing participation from scholars and institutions outside the West. Their analysis also identified the major topics and research power dynamics shaping the field [20]. In Korea, Han and Kim identified four major themes: AI literacy and curriculum development in schools, language education supported by chatbots, convergence talent programs using AI and big data, and human-AI collaboration [24]. Kim and Sohn found that studies of generative AI focused mainly on writing, language education, and programming. Learner experience and teaching methods were the most common themes, qualitative studies were more frequent than quantitative studies, and the studies reported more positive than negative attitudes toward generative AI [25].

These studies differ in emphasis. Research based on global databases mainly describes the growth of the field, its leading contributors, collaboration patterns, and overall research structure, whereas Korean studies examine domestic themes and specific educational applications in greater detail. Existing studies, however, have generally focused on Korean domestic literature, AI education as a broader field, or the post-ChatGPT phase alone. Parallel and longitudinal comparison of Korean and Chinese higher education AI research based on nationally indexed journal corpora remains limited. Recent bibliometric work on China’s higher education digital transformation further demonstrates the value of tracing long-term thematic change in domestic corpora, especially in non-English-speaking contexts [26]. Taken together, prior studies point to the need for a research design that is sensitive to database coverage, language, indexing practices, and national scholarly context.

Korean and Chinese higher education AI research circulates partly through different scholarly infrastructures. The Korean corpus examined in this study consists of KCI-indexed journal literature, while the Chinese corpus consists of CSSCI-indexed journal literature. Given these differences, the present study treats the two corpora as distinct bodies of domestic literature rather than as a single comparable dataset. This framing avoids treating publication counts as direct indicators of national research productivity and keeps the focus on how AI-related themes are organized within each corpus.

By connecting bibliometric mapping with a comparative understanding of Korean and Chinese research contexts, the study offers a context-sensitive account of how AI, generative AI, and related educational concerns have developed and been organized within the two scholarly corpora.

3.1 Research Design

This study employed a comparative bibliometric design to examine higher education AI research in Korea and China from 2016 to 2025, an approach suited to mapping the structure and development of fast-growing research fields [22]. CiteSpace was selected as the main analytical tool because it supports science mapping, cluster detection, timeline visualization, centrality analysis, and burst detection in scientific literature [23, 28].

The study compared two nationally indexed journal corpora rather than merging them into a single multilingual network. The Korean corpus consisted of journal articles indexed in the Korea Citation Index (KCI), and the Chinese corpus consisted of journal articles indexed in the Chinese Social Sciences Citation Index (CSSCI). Separate analysis was necessary because the two corpora differ in language, indexing practices, database coverage, and keyword convention. The comparison therefore focused on thematic structure, temporal pattern, and research orientation, rather than on direct node equivalence between the English (KCI) and Chinese (CSSCI) keyword terms used in the analysis.

3.2 Data Collection and Screening

The search strategy was designed to identify journal articles in which AI or generative AI was a substantive object of study within higher education. The main search combined higher education terms with general AI terms, while the supplementary search combined higher education terms with generative AI terms for the 2023-2025 period. Search terms were adapted to the metadata structure and language conventions of each corpus. Korean-language queries were used for KCI and Chinese-language queries were used for CSSCI, with commonly used English abbreviations and names retained where applicable. English translations of the search terms are presented in Table 1.

Table 1.

Search Strategy and Data Source Criteria

The searches were executed in the original languages of each corpus. The KCI main query was “(고등교육 OR 대학 OR 대학교 OR 대학생) AND 인공지능,” and the supplementary query was “(생성형 OR ChatGPT OR 챗GPT OR LLM OR ‘대규모 언어모델’) AND (대학생 OR 고등교육 OR 예비교사).” The CSSCI main query was “(高等教育 OR 高校 OR 大学) AND 人工智能,” and the supplementary query was “(高等教育 OR 高校 OR 大学) AND (生成式人工智能 OR ChatGPT).”

Retrieval, duplicate removal, and screening were reported through a flow diagram informed by PRISMA 2020 reporting principles [27]. Records were included when AI or generative AI was examined within a higher education context, including university students, faculty, teacher education in higher education contexts, curriculum, assessment, university services, and governance. Records were excluded when they focused only on K-12 or early childhood education without university relevance, treated AI only as a peripheral background term, consisted of non-research material, or were records incorrectly matched to higher education. After duplicate removal and eligibility screening, the final analytic corpora were fixed at 923 KCI-indexed articles and 1,037 CSSCI-indexed articles (Figure 1).

Figure 1.

Flow Diagram of Data Collection and Screening for the KCI and CSSCI Corpora

Each included record was also classified according to whether it was retrieved by the main search, the supplementary search, or both. In the KCI corpus, 332 of the 806 records published during 2023–2025 were retrieved only through the supplementary search. These comprised 33 records in 2023, 146 in 2024, and 153 in 2025. In the CSSCI corpus, 69 of the 719 records published during the same period were retrieved only through the supplementary search. These comprised 20 records in 2023, 19 in 2024, and 30 in 2025.

3.3 Keyword Field Selection and Preprocessing

Metadata conventions differed between the two corpora, so keyword coverage was examined before analysis. The same selection principle was applied to both corpora: the keyword field with the highest coverage and the most consistent structure was selected for analysis. In the KCI corpus, English author keywords were available for 915 of 923 records (99.1%), whereas Korean keywords covered 788 records (85.4%). In the CSSCI corpus, Chinese author keywords were available for 1,007 of 1,037 records (97.1%), whereas usable English keywords were available for 406 records (39.2%). On this basis, the KCI corpus was analyzed using English author keywords, and the CSSCI corpus was analyzed using Chinese author keywords (Table 2).

Table 2.

Keyword Field Coverage and Preprocessing Summary

English author keywords were selected for the KCI corpus because of their higher coverage (99.1% vs. 85.4% for Korean keywords), whereas Chinese author keywords were the only viable field for the CSSCI corpus, given the limited coverage of usable English keywords (39.2%). This data-driven choice means the two networks were built in different languages; the analysis therefore compared their thematic structure, temporal patterns, and research orientation rather than seeking direct equivalence between individual English and Chinese terms. English translations of Chinese keywords were used only for reporting and interpretation, not as analytical inputs.

Keyword normalization was conducted separately within each corpus. Spelling and capitalization variants, abbreviations and full forms, singular and plural forms, and clearly synonymous expressions were standardized. For example, generative artificial intelligence and generative AI, pre-service teacher and pre-service teachers, and large language model and large language models were merged within the KCI keyword field. Broad boundary terms, including artificial intelligence, AI, higher education, university, and their Chinese equivalents, were excluded only when they functioned as general retrieval terms rather than specific research themes. This procedure affected only the keyword field; no bibliographic records were deleted. After preprocessing, the KCI corpus retained 2,434 unique analytical keywords, and the CSSCI corpus retained 2,184 unique analytical keywords (Table 2).

3.4 Bibliometric Analysis and Comparative Strategy

The study combined descriptive bibliometric analysis with knowledge mapping in CiteSpace 6.4.R1 [23, 28]. Descriptive statistics were used to calculate annual publication counts and the distribution of records across the three analytical phases. CiteSpace was used to generate keyword co-occurrence networks, betweenness centrality values, log-likelihood ratio (LLR) cluster labels, timeline views, and keyword burst results. In the visualized networks, nodes represent keywords and links represent co-occurrence relationships. Node size indicates keyword frequency, and link strength indicates the closeness of co-occurrence. Betweenness centrality was used to identify keywords that connect otherwise separate thematic areas [29, 30]. Modularity Q and silhouette S values were used to assess the overall structure and internal consistency of the clusters [31, 32]. The mean year of each cluster was used to locate its position in the research timeline. Burst detection was used to identify keywords that showed a marked increase during a specific period [33].

The same CiteSpace settings were applied to the KCI and CSSCI corpora. The analysis used one-year time slices from 2016 to 2025, keyword nodes, and the g-index selection criterion with k = 25. The displayed CiteSpace parameters were LRF = 3.0, L/N = 10, LBY = 5, and e = 1.0. Cluster labels were generated using the LLR algorithm, and keyword bursts were detected using the CiteSpace burst detection function. No pruning algorithm was applied to the displayed networks.

The burst results were treated as supplementary evidence. Because the 2023–2025 period is relatively short, burst detection may be sensitive to parameter settings and may not fully capture newly emerging generative AI terms. For this reason, the interpretation of the 2023–2025 phase was based primarily on publication growth, keyword frequency, betweenness centrality, cluster structure, and timeline evolution, with burst results used as supporting evidence. To reduce possible frequency bias introduced by the supplementary search, the term “generative AI” was not interpreted on the basis of frequency alone. Its role was examined together with its centrality, cluster membership, timeline position, and co-occurrence with related terms such as writing education, large language models, prompt engineering, AI literacy, and academic integrity.

To examine whether the unequal distribution of records between 2016-2022 and 2023-2025 affected the keyword results, each corpus was divided into two period-specific sub-corpora. Using the same keyword normalization and exclusion criteria as the main analysis, keyword frequency and betweenness-centrality rankings were calculated and compared within each period.

The comparative analysis proceeded in three stages. First, annual publication trends were examined to identify how research activity developed across the three analytical phases: the pre-pandemic AI phase, the pandemic-driven digital transformation phase, and the generative AI phase. Second, keyword co-occurrence networks, centrality values, and LLR clusters were used to examine the thematic structure of each corpus. Third, timeline views and keyword burst results were used to identify topics that became more visible after 2023. The final comparison focused on how the two corpora organized AI-related higher education research, rather than on direct comparison of national publication volume.

4.1 Publication Trends Across 2016–2025

The first research question examined how AI research in higher education developed in Korea and China from 2016 to 2025. Publication trends show a common expansion after 2023, but the timing and concentration of growth differed between the two corpora (Fig. 2). In the KCI corpus, 19 records (2.1%) appeared in the pre-pandemic AI phase, followed by 98 records (10.6%) during the pandemic-driven digital transformation phase. The largest share of the corpus was concentrated in the generative AI phase, with 806 of 923 records (87.3%) published between 2023 and 2025. This concentration reflects both publication growth and the design of the supplementary search: 332 of these 806 records were retrieved only by the supplementary search. Among the 591 records in the full KCI corpus retrievable through the main search, the share published between 2023 and 2025 remained high at 80.2%. The CSSCI corpus followed a more gradual pattern of accumulation, with 97 records (9.4%) in the first phase, 221 records (21.3%) in the second phase, and 719 records (69.3%) in the third phase. Among these records, 69 were retrieved only by the supplementary search. Of the 968 records retrievable through the main search, 67.1% were published between 2023 and 2025, a proportion close to that of the full CSSCI corpus (Table 3).

Figure 2.

Annual Publication Trends in the KCI and CSSCI Corpora (2016–2025)

Table 3.

Publication Distribution by Analytical Phase

The phase distributions indicate both a shared post-2023 expansion and a difference in temporal formation. In both corpora, the increase after 2023 suggests that generative AI became a major focus of scholarly attention. The KCI corpus, however, was more concentrated in the final phase, with almost nine in ten records appearing after 2023. By contrast, the CSSCI corpus had already accumulated a larger body of AI research in higher education before the generative AI phase. The comparison therefore shows not only growth in publication volume, but also different temporal patterns: an expansion concentrated within a short period after the release of ChatGPT in the KCI corpus and a longer trajectory of accumulation followed by further growth in the CSSCI corpus. This contrast provides the basis for examining whether the two corpora also differ in keyword structures, thematic clusters, and emerging topics.

4.2 Keyword Structure and Thematic Clusters

The second research question examined the thematic structures and keyword clusters of the KCI and CSSCI corpora. Table 4 presents selected keywords with high frequency or high betweenness centrality in each corpus. These keywords are interpreted together with the co-occurrence networks and LLR clusters to identify the thematic organization of each literature.

Table 4.

Selected High-Frequency and High-Centrality Keywords by Corpus

4.2.1 Keyword Structure of the KCI Corpus

In the KCI corpus, generative AI was the most frequent keyword and also showed the highest betweenness centrality, with 241 occurrences and a centrality value of 0.79 (Table 4). This indicates that generative AI functioned not only as a frequently occurring term but also as a connecting node across different parts of the keyword network. Other keywords with notable centrality included AI education (0.25), pre-service teachers (0.20), liberal arts education (0.12), and university education (0.10). The co-occurrence network (Fig. 3) places generative AI at the center of a closely connected set of terms, including AI education, pre-service teachers, AI literacy, writing education, computational thinking, digital literacy, and large language models. Overall, the KCI network is organized around generative AI, teaching practice, teacher preparation, learner capability, and literacy-related concerns.

Figure 3.

Keyword Co-occurrence Network of the KCI Corpus

The LLR cluster analysis further clarifies the thematic organization of the KCI corpus. Table 5 reports 13 interpretable LLR clusters from the KCI keyword network, with cluster labels following the original CiteSpace numbering. The cluster solution showed a modularity Q of 0.62 and a mean silhouette S of 0.89, indicating a clear cluster structure and a high level of internal consistency (Fig. 4; Table 5). The largest cluster, #0 generative AI, had a mean year of 2023 and included terms such as generative AI, AI education, design education, instructional design, and prompt engineering. This cluster confirms that generative AI occupied a central position in the recent KCI keyword structure.

Table 5.

LLR Keyword Clusters in the KCI Corpus

Figure 4.

LLR Keyword Cluster Map of the KCI Corpus

The remaining clusters can be grouped into several educational domains. Teacher-related themes appeared in #3 pre-service teachers and #10 teacher education, while curriculum-oriented themes appeared in #2 AI convergence education and #4 fourth industrial revolution. Other clusters reflected writing and language education (#5 writing education), concerns related to thinking and literacy (#1 critical thinking and #6 computational thinking), technical methods (#7 machine learning), and ethical issues (#8 AI ethics). More recent clusters, including #9 large language models, #12 AI image generator, and #14 student perception, had mean years in 2023 or 2024. Overall, the cluster results suggest a temporal movement from earlier themes related to teacher education, machine learning, and curriculum development toward post-2023 concerns with generative AI, large language models, critical thinking, AI image generation, and student perception.

4.2.2 Keyword Structure of the CSSCI Corpus

In the CSSCI corpus, the keyword structure was organized around a different set of anchors. Talent cultivation showed the highest frequency and betweenness centrality, with 69 occurrences and a centrality value of 0.40 (Table 4). Smart education also had a high centrality value (0.36), followed by AI literacy (0.14), digital literacy (0.12), big data (0.09), and university faculty (0.09). The co-occurrence network (Fig. 5) therefore shows a structure centered on talent cultivation and smart education, with links to digital literacy, big data, faculty development, information services, and digital transformation. Compared with the KCI corpus, the CSSCI keyword network appears less concentrated around a single dominant node and more oriented toward higher education reform, institutional development, and capability formation.

Figure 5.

Keyword Co-occurrence Network of the CSSCI Corpus

The LLR cluster results further specify this structure. The CSSCI corpus formed 11 clusters, with a modularity Q of 0.75 and a mean silhouette S of 0.94, indicating a clear and internally consistent cluster structure (Fig. 6; Table 6). The largest cluster, #0 talent cultivation, had a mean year of 2020 and included terms related to talent cultivation, industry-education integration, emerging engineering education, discipline construction (学科建设, the strategic development of academic disciplines), and academic program development (专业建设, the development of undergraduate majors). This cluster suggests that AI research in higher education in the CSSCI corpus was closely connected to talent development and structural reform in higher education.

Figure 6.

LLR Keyword Cluster Map of the CSSCI Corpus

Table 6.

LLR Keyword Clusters in the CSSCI Corpus

Other clusters reflected several related domains. Faculty and knowledge production appeared in #1 university faculty, while #2 smart education included terms related to smart education, adaptive learning, and digitalization. Cluster #3 AI literacy connected AI literacy, information literacy, university libraries, and academic integrity. Technology-enabled services and applicationoriented themes appeared in #4 smart services and #5 application scenarios. Additional clusters addressed interdisciplinarity (#6 interdisciplinarity), big data and educational technology (#7 big data), the Double First-Class initiative (双一流, the national initiative to build world-class universities and disciplines) (#8 Double First-Class), human-centered data governance (#9 human-centered governance), and industry-university-research collaboration (产学研用) (#10 industry university research application). The mean years show that earlier clusters were associated with big data, the Double First-Class initiative, and university-industry collaboration, whereas more recent clusters were related to AI literacy, application scenarios, and human-centered governance.

4.3 Thematic Evolution After 2023

The third research question examined which topics became prominent after 2023 and how they were connected to earlier thematic trajectories. The timeline views indicate that the KCI and CSSCI corpora entered the generative AI period from different prior structures (Figs. 7 and 8). In the KCI corpus, earlier themes were associated with teacher education, machine learning, the fourth industrial revolution, liberal arts education, software education, and computational thinking. During the 2020–2022 period, AI convergence education, AI ethics, pre-service teachers, and AI literacy became more visible. The cluster and timeline results show increasing visibility after 2023 for generative AI, large language models, AI image generators, student perception, writing education, machine translation, and critical thinking. The KCI trajectory therefore suggests a movement from broader AI education and teacher education concerns toward generative AI-related teaching, writing, learner experience, and academic practice.

Figure 7.

Timeline View of the KCI Corpus

Figure 8.

Timeline View of the CSSCI Corpus

The CSSCI corpus showed a longer developmental path before the generative AI period. Earlier and middle-period themes included big data, Double First-Class, emerging engineering education, industry-education integration, smart education, talent cultivation, and university faculty. The cluster and timeline results show increasing visibility after 2023 for AI literacy, application scenarios, digital intelligence empowerment, interdisciplinarity, data governance, and human-centered governance. Rather than forming a timeline centered on one generative AI keyword, the CSSCI corpus incorporated recent AI-related concerns into an existing discussion of higher education reform, smart education, talent development, discipline construction, and institutional governance.

The keyword burst results were used as supplementary evidence because the 2023–2025 period is short and burst detection can be sensitive to parameter settings. In the KCI corpus, earlier burst terms included future education, computational thinking, the fourth industrial revolution, software education, AI education, subject education, AI ethics, machine learning, and convergence education (Fig. 9). These terms indicate that the pre-2023 burst structure was mainly related to future-oriented education, AI education, software education, computational thinking, and broader technology education. After 2023, the burst terms shifted to writing education and machine translation, suggesting that the most recent burst signals were linked more closely to language, writing, and academic practice.

Figure 9.

Keyword Burst Results in the KCI Corpus

The CSSCI keyword burst results also support the temporal pattern observed in the timeline view (Fig. 10). Earlier burst terms included transformation, emerging engineering education, Double First-Class, big data, smart education, and academic program development. Later burst signals shifted toward metaverse-related discussion and digitalization. These results suggest that the CSSCI corpus did not simply shift from general AI to generative AI. Instead, recent AI-related topics were added to an already established reform and digital transformation trajectory.

Figure 10.

Keyword Burst Results in the CSSCI Corpu

Across the timeline and burst evidence, the two corpora show two different temporal patterns. In the KCI corpus, the most recent period shows a rapid shift toward generative AI, writing, language use, learner perception, and academic practice. In the CSSCI corpus, the post-2023 period appears as an extension of a longer trajectory that had already connected AI with smart education, talent cultivation, disciplinary development, and institutional reform. For this reason, the interpretation of emerging topics after 2023 relies mainly on publication growth, keyword centrality, cluster structure, and timeline evolution, while burst results are used only as supporting evidence.

The supplementary period-split analysis provided a further check on these temporal patterns. In the KCI corpus, AI education occupied the leading frequency and centrality positions during 2016–2022, alongside pre-service teachers and liberal arts education. During 2023–2025, generative AI held the leading position in both frequency and betweenness centrality, well ahead of the remaining keywords, while AI literacy, university students, and writing education became more visible. In the CSSCI corpus, talent cultivation, smart education, and big data anchored the earlier period. During 2023–2025, terms related to generative AI, including large language models, rose in prominence, while talent cultivation remained among the leading terms in both frequency and centrality. Several keywords occupied leading centrality ranks in the CSSCI sub-corpus, whereas generative AI ranked first in the later KCI sub-corpus. Although differences in corpus size may affect the absolute frequency and centrality values of individual keywords, the period-specific rankings show that the broader thematic transition remained visible in both corpora.