Artificial Intelligence Literacy in Higher Education (2020–2026): A Bibliometric Scoping Review of a Rapidly Emerging Literature

1. Introduction

Artificial intelligence literacy, the competencies needed to understand, use, evaluate, and critically engage with AI systems, has shifted from a specialist interest to a central concern across higher education. The release of widely accessible generative AI tools in late 2022 sharply intensified this attention, raising questions about curriculum, assessment, academic integrity, and graduate preparedness. As the literature has grown, its scope has become difficult to survey by narrative means alone.

This review quantitatively maps the higher-education AI-literacy literature. Rather than appraising individual studies, we characterize the field’s size, growth trajectory, thematic composition, disciplinary spread, and geographic distribution using reproducible, automated methods applied to a large bibliographic corpus. We frame the work as a bibliometric scoping review, following PRISMA-ScR reporting guidance [16].

2. Methods

2.1. Protocol and Registration

The review followed a bibliometric scoping-review design and is reported in accordance with PRISMA-ScR [16]; a completed checklist is provided (S2 File). This higher-education review was not separately registered; it uses the same protocol and methods as a companion medical/health-professions review, whose protocol was archived retrospectively on the Open Science Framework. This is a descriptive mapping review; formal risk-of-bias appraisal was not conducted, consistent with scoping-review methodology.

2.2. Data Source and Search

We used OpenAlex as the sole bibliographic source. Works were retrieved if their title or abstract matched an exact-phrase AI-literacy family (“AI literacy”, “artificial intelligence literacy”, “generative AI literacy”, “GenAI literacy”, “algorithmic literacy”, “AI competency/competencies”) and were published on or after 1 January 2020. Retrieval used cursor pagination; the parent corpus comprised 12,326 deduplicated works. Verbatim query strings and the full filter rule are provided in the supplementary methods (S1 File).

2.3. Higher-Education Eligibility (Strong-Signal Filter)

From the parent corpus, we selected higher-education works using a strong-signal rule: a higher-education term (e.g., “higher education”, “university student”, “undergraduate”, “postgraduate”, “faculty member”) in the title, or at least two distinct such terms in the title and abstract. This yielded 1,910 works. The rule was designed to exclude papers that mention a higher-education term only incidentally.

2.4. De-Duplication, Screening, and Text Preparation

Records indexed across multiple sources were deduplicated using a normalized title key (removing index-source parentheticals), removing 197 near-duplicate records. Works with a usable English abstract (>100 characters) were retained for clustering (n = 1,483); records without were described but not clustered (Figure 5). Screening was automated and rule-based rather than conducted by independent human reviewers, a limitation discussed in Section 4.

2.5. Thematic Clustering

Titles (up-weighted) and abstracts were vectorized with sublinear TF-IDF (unigrams + bigrams, domain stop-words removed, ≤8,000 features) and then reduced by truncated SVD to 80 components (variance retained ≈0.201). k-means (k = 8) grouped the documents. Silhouette values were uniformly low across k = 6–12 (≈0.03–0.05), as expected for short, topically overlapping text; k was chosen for interpretability rather than for the silhouette maximum. One cluster consisted of Korean-language records grouped by language rather than by topic and was excluded, leaving seven substantive themes. Two reviewers labeled each theme based on its top terms and representative works. Full clustering parameters and stability diagnostics are reported in the supplementary methods (S1 File).

2.6. Cluster Stability

Stability was quantified by re-running k-means with 10 random initializations and computing the adjusted Rand index (ARI) between partitions. The mean ARI (seed-0 reference) was 0.534 (min 0.446, max 0.628); averaged over all 45 initializations pairs, it was 0.552 (range 0.332–0.813). Both values indicate that the broad thematic structure is reproducible, whereas per-paper assignments are borderline. Themes are therefore presented as an interpretive lens, not a definitive taxonomy.

2.7. Ethics

The review relied solely on aggregated published bibliographic metadata and did not involve human participants; therefore, ethical approval was not required.

3. Results

3.1. Growth

Annual output remained in single digits through 2022, then rose steeply: 27 works in 2023, 157 in 2024, and 750 in 2025; the first half of 2026 already contains 958 records (Figure 1). Overall, 89.4% of the corpus appeared in 2025–2026.

3.2. Thematic Structure

Clustering yielded seven themes (Table 1; Figure 2 and Figure 4). The two largest concerns are AI-literacy courses and curriculum design (n=342), exemplified by early curriculum-integration models [1,5] and course evaluations [12], and institutional integration, policy, and academic integrity (n=306), which includes the most-cited works on the implications of generative AI for higher-education practice [2,3,4,9]. Student surveys of usage, attitudes, and scale development [7,10,14], and behavioral intention/acceptance modeling [6], form a substantial methodological block, alongside faculty-focused studies of AI use and self-efficacy [13]. The fastest-growing themes are systematic/scoping reviews and qualitative studies of writing and AI-literacy components (both >18× 2025–26 vs 2023–24), signaling a field beginning to consolidate and probe mechanisms qualitatively. Classroom-embedded studies, for example, generative AI as a virtual tutor [8], and pedagogically framed analyses [15] recur across themes, and cross-disciplinary contributions extend into the health professions [11].

3.3. Disciplinary Spread

Disciplinary tags (non-exclusive; Figure 3) show concentrations in the humanities/languages and STEM/computing, with smaller but distinct contributions from business/economics and health professions. This distribution reflects both writing-focused concerns and technical-curriculum efforts.

Figure 3. Disciplinary distribution (non-exclusive tags).

Figure 3. Disciplinary distribution (non-exclusive tags).

Figure 4. Theme sizes for the seven substantive themes.

Figure 4. Theme sizes for the seven substantive themes.

Figure 5. PRISMA-ScR flow of record identification, screening, eligibility, and inclusion.

Figure 5. PRISMA-ScR flow of record identification, screening, eligibility, and inclusion.

3.4. Geographic Distribution

Author affiliations were led by the United States, China, and Indonesia, followed by India, the United Kingdom, and Malaysia (Table 3), indicating a globally distributed yet unevenly concentrated literature.

Table 2. Disciplinary tags (non-exclusive, n = 1,910).

Discipline	Papers
Humanities / language	690
STEM / computing	204
Business / economics	180
Health-professions	98
Social sciences / education	77
Law	33
Library / info science	26

Table 2. Disciplinary tags (non-exclusive, n = 1,910).

Discipline	Papers
Humanities / language	690
STEM / computing	204
Business / economics	180
Health-professions	98
Social sciences / education	77
Law	33
Library / info science	26

Table 3. Leading countries by author affiliation (non-exclusive).

Country	Papers
US	237
CN	182
ID	105
IN	79
GB	71
MY	60
TR	55
AU	49

Table 3. Leading countries by author affiliation (non-exclusive).

Country	Papers
US	237
CN	182
ID	105
IN	79
GB	71
MY	60
TR	55
AU	49

4. Discussion

4.1. Principal Findings

AI literacy research in higher education has expanded from a marginal topic to a substantial, fast-growing literature, 89.4% of which has been published since 2025. Its center of gravity lies in curriculum design and institutional policy, with a large methodological investment in surveys and acceptance models, and an emerging qualitative and review literature.

4.2. Interpretation

The dominance of curriculum design and policy themes suggests the field is oriented toward practical implementation. The rapid rise in reviews indicates early consolidation, while the growth of qualitative, writing-focused work reflects concern with academic integrity and the effects of generative tools on student writing.

4.3. Limitations

Three limitations warrant emphasis. First, the corpus derives from a single database (OpenAlex); although its coverage is broad and includes gray literature, single-source retrieval may underrepresent some venues. Second, screening and extraction were automated and rule-based, without independent human dual-screening. Third, clusters are moderately stable (ARI ≈0.534); themes serve as an interpretive lens rather than a definitive taxonomy. Growth ratios use add-one smoothing and a partial 2026 year, and citation-based prominence is subject to citation lag.

4.4. Conclusions

The higher-education AI-literacy literature is young, rapidly expanding, and organized around curriculum, policy, and measurement. Priorities for the field include longitudinal and intervention studies, validated cross-context instruments, and broader geographic representation.