Adoption of artificial intelligence in academic libraries: A systematic review of current practices, challenges, and research opportunities

Synthesis: Ayinde, Ebiefung and Oladokun review 29 studies published between 2020 and 2024 to ask what academic libraries actually do with AI, what stops them, and what should be studied next. The practices sort into six dimensions — technical services, reference and information services, administrative and circulation services, collection development, information literacy and user education, and professional development — and reference is the center of gravity: AI for directional and ready-reference questions appears in 34.48% of the reviewed studies. The barriers are organizational more than technical. Funding is named in 48.28% of studies and technical skill and data-quality gaps in 58.62%, while fear of job loss appears in only 13.79%. The authors read the pattern as a field automating transactional work first and leaving judgment work human, and they call for user-driven research rather than more pilot descriptions.

Key Findings

1. The review follows PRISMA across Web of Science, Scopus and LISA, moving from 319 records to 221 screened, 80 full texts and 29 included studies; 28 are journal articles and one is a conference paper, and none used a mixed-method design.
2. Reference and information services are the most prominent application: AI for answering directional and ready-reference questions appears in 34.48% (9) of the 29 studies.
3. Cataloguing and chat-transcribed or expedited reference tie for second at 27.58% (8) each; classification appears in 24.13%, indexing in 10.34%, and intelligent content extraction, acquisition assistance, metadata generation, automation and rare-resource curation in a single study each (3.45%).
4. Funding is the most cited institutional barrier at 48.28% (14 studies).
5. Technical challenges — gaps in technological skill, hallucinated or confusing bot responses, insufficient datasets, deficient bibliographic data — appear in 17 of 29 studies (58.62%).
6. Infrastructure problems are cited in 12 studies (41.38%), including inadequate infrastructure (20.69%), limited power supply (10.34%), poor networking and outdated technology.
7. Ethical concerns — academic integrity, bias and false information, intellectual property, plagiarism, copyright, privacy, cybersecurity — appear in 31.03% of studies.
8. Fear of job loss is the most cited individual challenge, in 4 of 29 studies (13.79%), with resistance to change in 6.89%.
9. Environmental challenges appear in four articles (13.8%): public misunderstanding of AI capability, global governance disparities, and exposure to international standards.
10. Only one study used a quality checklist (Liu and Liu, 2023) with an inclusion threshold of 5 or higher; 15 of the 29 (51.72%) were published in 2024, and the Journal of Academic Librarianship and Library Philosophy and Practice each supplied three (10.3%).

What the adoption pattern says

The distribution is the finding. The library functions that AI has entered most often are the ones with a bounded request and a checkable output: a directional question, a catalogue record, a transcribed chat. The functions that require deciding what a source means, or what a collection should be, remain human in this literature — and the reviews of instruction elsewhere in the knowledge base find the same boundary, with librarians consulted when a recommender is least confident.

That makes the funding result less decisive than it looks. If the binding constraint were only money, the applications would cluster where a purchase buys the most capacity. Instead the studies show a field buying reference automation while reporting skill gaps, data-quality problems and policy absence in the same breath. The authors' framing — six practice dimensions crossed with a technology-organization-environment reading of barriers — is a reasonable way to hold those together, but it also shows that the organizational variables are not a residual: they are where most of the reported difficulty sits.

A literature that is mostly description

Twenty-nine studies is a thin base for a professional field, and the composition is thinner than the count. None used a mixed-method design, the quality appraisal is applied once, and the modal contribution is a survey of one institution's use or intentions. The authors say so directly, calling for more user-driven research on applications and for retraining that covers chatbot integration and prompt engineering rather than generic AI awareness.

For a reader trying to act, that changes the status of every percentage above. The figures describe what has been published, not what libraries do; a practice that no one has surveyed is invisible here. The review is best used as a map of where the literature has looked, and as an argument that the next study should measure a service rather than ask about attitudes.

What this means for practice

• Choose one service and instrument it. The literature is full of adoption surveys and short of service evaluations; a library that measures a reference chatbot against the questions it was meant to answer contributes more than another attitude survey.
• Budget for staff skill, not only licenses. Funding is the headline barrier, but skill and data quality appear in more studies; a license without a training plan reproduces the gap the review documents.
• Write the policy before the pilot. Ethical concerns appear in nearly a third of studies, and policy absence is a recurring organizational barrier; the pilot's data-handling and integrity rules should exist before the first patron uses it.
• Keep the human escalation path. The applications that dominate are transactional, which is also where automation is safest; the judgment questions should keep a route to a librarian.

Limitations

• The review covers 29 English-language studies from 2020–2024 across three databases; work in other languages, other databases and books or chapters is excluded, and the authors note that book chapters may hold foundational theory.
• No study in the sample used a mixed-method design, so the reviewed evidence is dominated by single-method surveys and descriptions rather than triangulated service evaluations.
• Quality appraisal against a five-area checklist with a threshold of 5/10 filters weak studies but cannot create outcome data the original studies did not collect.
• The authors describe their own method as conceptual and note that no primary data were collected; the six dimensions and the barrier categories are analytical groupings, not validated constructs.
• Rapid change in the technology means the 2020–2024 window under-represents current tools, particularly agentic and retrieval-augmented systems.

Connected Concepts

• Libraries
• Information and Media Literacy
• Librarians
• AI Literacy
• Generative AI
• AI Governance

Citation

Ayinde, L., Ebiefung, R., & Oladokun, B. D. (2026). Adoption of artificial intelligence in academic libraries: A systematic review of current practices, challenges, and research opportunities. The Journal of Academic Librarianship, 52, 103185.