Generative AI in Academic Libraries: A Systematic Review of AI-Mediated Reference Services, Librarian Roles, and User Trust

Generative artificial intelligence (AI), and in particular large language models in the form of ChatGPT, is gradually getting integrated into the systems and workflows of academic libraries. Consequently, they are changing the nature of reference services, the work of librarians, and the information-seeking behaviour of users. This systematic review aimed to explore the impact of generative AI on the quality of reference services, restructuring of librarian tasks, and trust of users in academic libraries. A systematic search in Scopus, and Science Direct was conducted to find the relevant literature. A total of 14 major empirical studies, published between 2020 to 2025 were selected and analysed that included experimental assessments, cross sectional survey, qualitative interviews, mixed-methods research and quasi experimental audit studies across different geographic locations. The findings revealed a significant performance of AI-mediated reference services but generative AI can only be applied to low-complexity questions. However, it demonstrated severe limitations in responding to complex research questions, contextual questions, and citation verification, as indicated by high rates of hallucinations and bibliographic errors. However, the user-trust was also conditional, depending on the perceived utility, transparency, and the institutionalisation of AI services. Thus, generative AI should be understood as a socio-technical system, in which the efficacy is determined by human control, governance systems, and correlation with the fundamental values, which are the foundation of academic libraries. Thus, this study emphasize instead of replacing librarians, generative AI restructured professional practice in favour of mediation, verification, ethical oversight, and the delivery of AI literacy education.

Keywords: Generative artificial intelligence, academic libraries, reference services, librarian roles, user trust, information integrity, chatgpt.

1. INTRODUCTION

Chatbots based on generative artificial intelligence (GenAI), specifically, large language models (LLMs), are swiftly transforming academic communities. It helps to discover information, evaluate evidence, and obtain help with conducting research and composing scholarly writing. Student uptake is now high enough to affect the demand profile for academic library services. The UK Higher Education Policy Institute (HEPI) (2025) student Generative AI survey 2025 reports that 88% of students reported using generative AI tools for assignments (up from 53% the previous year). Although GenAI is becoming part of the daily routines of academic research, academic libraries are under ever-increasing pressure to act across three interconnected areas: maintaining the quality of reference services, recalibrating labour roles and skills, and preserving users’ trust. Early experimental findings indicate that GenAI can offer daily advice but not the highest-quality reference results. In chronologically organised testing of ChatGPT across different types and complexity levels of reference queries, Lai (2023) found that the system performed poorly on advanced research questions and complex inquiries, including precisely the kinds of queries that characterise academic library reference work. One of the main weaknesses is the model’s tendency to produce plausible but incorrect statements, such as bibliographic errors. A socio technological perspective in library practise acknowledges that AI results are governed by organisational settings, staff processes, professionalism, and policy alignment within an institution. Empirical studies across disciplines by Mugaanyi et al. (2024) were cross-disciplinary and average recall had correct citations of outputs only approximately 60‛%. In the case of libraries, these findings present a threat whereby, provided the use of generative AI to answer reference queries, recommend sources, or teach citation practise, there is a possibility of such uncorrected inaccuracies. These inaccuracies may lead to poor service provision and loss of confidence in library directions as the users make mistakes. 

Simultaneously, there is also more effort to apply generative AI to the organisational level. According to the evidence presented by academic libraries, generative AI is being implemented in service delivery, user care, and internal processes. Indicatively, Gmiterek and Kotuwl (2025) used a survey and content analysis to identify the degree of generative-AI adoption by academic libraries at Polish state universities (a mixed-methods study). The workforce implication is also evident as Cox (2025) finds out that the survey on AI and the library profession in UK across the country, that the current use and perception of AI by librarians and information professionals is listed on the current positions of issues of concern rather than being an issue of the future the survey on AI and the library profession UK wide. User trust is often underestimated but is the third variable indicating the connexion between the quality of references and the workforce change. As Deschênes and McMahon (2024) reported on the use of generative-AI chatbots in academic research, students use it, but they do not consider the information produced by generative AI helpful according to the data of the surveys. As Grammes (2024) notes, perceived usability advantages are in a better position than human reference interactions. However, the comfort factor will also tend to affect trust and preference where there are problems of credibility. 

Although there is an increased amount of empirical studies of generative AI in academic libraries, the research remains sporadic, which is why a specific systematic review is necessary. The literature at hand analyses the independent factors of the AI usage. It, however, does not centre on the joint impact on library services. Indicatively, correctness of answers and validity of reference are the major measures of reference quality as applied by Lai (2023). However, these outcomes are not explicitly related to these measures of performance that are so critical to workforce redesign like models of supervision, timely mediation, workflows of quality-assurance retribution, or objective markers concerning user trust. Cox (2025) concentrated on professional expectations and views under a sample survey of librarians, however, the study lacked comprehensive data on workforce perception, measured performance of services, and patron confidence in academic libraries.  Deschênes and McMahon (2024) documented patterns of using and attitude toward generative AI but did not compare differences in trust between AI use within library-branded services and independent use of external AI tools by students.

The given review fills these gaps by incorporating the available evidence on both sides of the service-workforce-trust spectrum in the academic library, and mainly focusing on generative AI-based reference and research support.  This systematic review aims to summarise recent empirical evidence on the effects of generative AI on academic library services and professional practise. Notably, it explores implications on the quality of reference services such as accuracy, completeness, contextual relevance, and citation integrity, changes in librarian roles, competencies, and trends in user trust. The review will help illuminate evidence-based governance and the responsible incorporation of generative AI in academic libraries by providing comprehensive evidence.

2. Methods

2.1. Research Design

The current study used a systematic literature review (SLR) design framework to analyse the effects of generative artificial intelligence (GenAI) on the practice of providing reference services in academic libraries and the workforce, in particular, the quality of reference services, the role of staff, and user trust. The systematic literature review provides a comprehensive understanding of the issue by synthesising available evidence (Lame, 2019). This was a review in compliance with the established PRISMA principles of evidence-based information science protocols employed to minimise selection bias and provide methodological transparency. 

2.2. Search Strategy

The search strategy was aimed at ensuring that the best possible coverage of empirical evidence is given and at the same time ensure transparency, reproducibility and relevance to library and information science. The available evidence was found in two databases, one of which was Scopus and the other was ScienceDirect. These databases were chosen to provide balance between the depth of discipline and coverage of citation and the access to both the LIS-specific and interdisciplinary studies.   The selection of these databases was based on their comprehensive coverage of library and information science, information systems, and higher education research, as well as their exceptionally high indexing of peer-reviewed research journals. The searches of publications were restricted to from 2020 to 2025 in the topical area, as large language models have advanced rapidly and require the inclusion of reputable publications in the search field. This temporal boundary ensures that findings reflect the post-ChatGPT phase of generative AI development. Appropriate keywords with Boolean operators were used to determine available evidence.  Various search terms were used in all the databases that were chosen. Scopus included (ChatGPT OR ‘AI chatbot*’ OR ‘generative AI’) AND (‘reference librarian*’ OR ‘human reference’ OR ‘traditional reference’) AND (comparison OR comparative OR evaluation OR performance), leveraging Scopus’s structured indexing and quality control for comparative empirical studies. ScienceDirect included (‘generative AI’ OR ChatGPT OR ‘AI-assisted system’) AND (‘librarian’ OR ‘job role’). Results of the search of the databases are described in Appendix A. To guarantee a consistent interpretation and a similar methodological appraisal, only English publications were taken into account. It was found that duplicate records were being used and as a result, they were eliminated before the screening stage.

2.3. Study Selection Criteria

2.3.1 Inclusion Criteria

The studies that met a set of clearly specified eligibility conditions were considered in the review process. The studies had to be eligible and contain primary empirical results of either of the qualitative, mixed-methods, or quantitative research designs. Studies recorded that explicitly interacted with generative AI or large language model-based systems (ChatGPT or GenAI). The setting of the research also was to be a context that lay in academic libraries, academic information services or closely related, highly academic settings. Studies reported at least one of the core outcomes, reference or answer quality, citation accuracy, librarian or staff roles and competencies, or user trust and credibility were included. Only studies published between 2020 and 2025 in peer-reviewed journals were included to ensure both rigour and relevance.

2.3.2 Exclusion Criteria

Review articles, systematic reviews, conceptual papers, editorials, opinion pieces, and policy commentaries were excluded. Pure technical AI studies that did not have a direct relationship with the services of an academic library or the information work in general were excluded. Research studies that did not exhibit methodological transparency, empirical evidence, or connection to library services or workforce implications were excluded. To maintain a contextually and analytically consistent research process, studies based on K-12 education (education from kindergarten till 12th grade) alone and those based on corporate knowledge management or non-academic public library environments were excluded.

2.4. Data Extraction

A data-extraction framework was created to enable comparability between the studies that were included. The information extracted included bibliographic information (author, year, country), study design, sample characteristics, library service context, and outcome variables. Special care was given to reference quality (e.g., the accuracy, completeness, citation reliability), workforce (e.g., redistribution of work, skills demanded, professional identity), and user trust (e.g., perceived credibility, reliance, willingness to trust AI-mediated services) indicators. Critical findings were documented in a systematic way.

2.5. Data Synthesis and Analysis

The diversity of the study designs and outcome measures made it necessary to use the narrative and thematic synthesis methodologies. The primary categorization of studies was on the focus of analysis: quality of references, assessments on the workforce, or reliability to the users. The identification of cross-cutting themes later, allowed the analysis of cross-domain interactions (e.g., mediation of workforce and trust, or reference accuracy and professional role redesign). The quality of methodology was evaluated to determine the robustness of the evidence base and to establish research gaps that are long term.

2.6. Ethical Considerations

The proposed study is limited to publicly available secondary data. Therefore, no ethical approval was necessary. The credibility of the included studies was ensured by providing correct citations, documenting all procedures, and reporting the true findings. There was no interpretive bias, as the inclusion criteria were adhered to and analytical procedures were followed systematically throughout the review.

3. Results

3.1. Data Screening

The selection of studies was conducted in accordance with the PRISMA framework to ensure transparency, consistency, and methodological rigour across all screening steps (Fig 1), PRISMA. A preliminary search of two major bibliographic databases, Scopus (n = 257) and ScienceDirect (n = 320), yielded 577 results. Before formal screening, 40 duplicate records were removed. After this, 537 records were screened based on title and abstract relevancy. At this point, 12 non-English language publications and 120 articles not specifically relevant to the topic were excluded. Afterwards, 285 peer-reviewed journal articles, including books, editorial papers, conference papers, and records without a particular focus on the use of generative AI in academic library services, were excluded. 60 doctoral dissertations, secondary review sources, and case studies were excluded. Afterwards, 9 records that were unavailable because of the absence of open access or institutional archives restrictions were also excluded. Unavailability refers to the unavailable full-text articles that either were not available through the open access or institutional repositories at the time of screening. This limitation can lead to biasness on available journals which can philtre away relevant but restricted studies. In the end, the final systematic review included fourteen studies that met all the eligibility criteria, which represents a highly focused and strictly elaborated evidence base. Full-text retrieval outcomes and exclusion reasons are briefly listed in Appendix A.

3.2. Quality Assessment

Overall, the quality appraisal indicates that the included studies are of moderate to high methodological quality, with strengths such as clarity of objectives, suitability of the design, and reporting of results, and recurrent limitations. The research questions of the cross-sectional and descriptive studies by Choudhury and Shamszare (2023), Deschênes and McMahon (2024), Haris et al. (2025), Deschênes and McMahon (2024), Ismail et al. (2024), and the descriptive research studies of Del Castillo and Kelly (2024), as well as Elsayed and Abusharhah (2025), were appraised in the Critical App All of these studies were deemed to generate sound and clear findings, thus making them relevant to academic library environments (see Appendix C).  Nevertheless, recruitment approaches were often rated as being unclear, which was also due to the prevalence of convenience, volunteer, or snowball sampling methods and thereby representativeness. Chigwada and Pasipamire’s (2024) study also showed that there was a deficiency in the size of the sample and the depth of its analysis, which also means that the risks of sampling bias are high.

The appraised qualitative studies by Chen (2025) and Kim (2025) that were evaluated with the help of the CASP Qualitative Checklist had high levels of coherence in their methods, had a clear purpose, proper qualitative designs, thorough thematic analysis, and clear statements of findings. The main weakness of these studies was that there was no explicit discussion on the issue of researcher-participant relationships or reflexivity, which limited the overall evaluation of the possible interpretive bias. Most of the quality criteria for the qualitative, quantitative, and integrative elements were met in the mixed-methods study by Saeidnia et al. (2024), as assessed using the MMAT (2018). However, the representativeness of the participants and the control of confounding factors were limitations that reduced generalisability. The quasi-experimental study by Wang et al. (2025), which was assessed using the JBI checklist, demonstrated high internal validity, efficient outcome measurement, and appropriate statistical analysis. The unclear rating pertained only to the pre- and post-measurement and follow-up criteria, which were not the focus of the audit-based study. Hence, the quality analysis justifies the inclusion of all studies but emphasises the need for careful interpretation, especially regarding generalisation and long-term institutional impact.

Importantly, no studies were excluded based on quality appraisal criteria, as all met the lowest methodological acceptability standards established by CASP, MMAT, or JBI. In turn, quality assessment was used to give greater weight to the interpretation of findings rather than to establishing eligibility. Experimental and quasi-experimental studies were accorded greater evidentiary weight in assessing reference accuracy and bias than survey-based and qualitative studies, which were more cautious in extrapolating to workforce preparedness or user confidence when the study context was not replicated. The reference to limitations identified, especially the non-probability sampling method, representativeness, and the unavailability of longitudinal designs, helped to understand generalisability constraints. To that end, the synthesis of results focuses on convergent patterns across methods and geographical locations, rather than on single-study influences, so that the conclusion reflects the strengths and consistency of the relationships within the evidence base rather than a single high-impact report. Quality assessments for each study, as per their designated appraisal tool, along with an overall summary of quality assessments, are provided in Appendices B and C.

3.3. Study Characteristics

The articles included in this review have a diverse methodology and analyse the field of generative AI and its use in the environment of academic library service, workforce practise, and user trust. The study designs of the fourteen major researches include experiment studies, big surveys, mixed-method studies, and the case study of qualitative analysis, thus demonstrating the complexity of using AI in libraries (see Table 1). The quality of references and reliability are also the subject of a significant percentage of the studies, especially through comparative indicators of the performance of ChatGPT compared to its actual or simulated reference queries. Empirical evidence shows that simple, factual, and navigational queries can be sufficiently met with generative AI, but systematic weaknesses with complex academic sources (advanced research queries, citation generations) are only identified by Lai (2023) and Mugaanyi et al. (2024).

The articles incorporated show the high professional interest but show a gap in the preparedness of librarians. In various parts of the world, the studies by Gmiterek and Kotuawa (2025), Elsayed (2025), Abusharhah (2025), Hussain (2024), and Kim (2025) through the case studies and surveys point at significant optimism on the prospects of AI. However, there are still long-term obstacles, such as the lack of skills, the shortage of governance, and the need to be ready to handle ethical issues. Together, such works highlight generative AI as a disruptive phenomenon, which alters the role of a librarian towards mediation, training AI literacy, quality control, and ethical accountability instead of eliminating it. The paradox of usage-trust observed in user-related studies that Deschênes and McMahon (2024) and Choudhury and Shamszare (2023) indicate may be due to high adoption rates, but low trust in AI-generated information on the quality of the output, transparency, and source credibility is also relevant. Interestingly, the experimental results used by Wang et al. (2025) are quite enticing because the findings can be generalised to the idea that contemporary large-scale language models can provide consistent reference answers to various demographic groups, which implies a means to reduce prejudice in the course of time through proper controls and model selection. The characteristics of the present study support the idea that the impact of generative AI in academic libraries is not a categorical phenomenon and depends on the human control versus the ability to provide services and the services based on the trust.

4. Findings

The results section is explicitly structured around a socio-technical integration model in which reference service quality outcomes shape user trust and, in turn, necessitate workforce mediation and governance responses. This model provides the analytical framework through which empirical findings are organised, interpreted, and synthesised across studies, ensuring that quality, labour, and trust are examined as interdependent rather than isolated dimensions.

4.1. Scope and Distribution of the Evidence Base

The review is anchored on fifteen main empirical studies, published in 2023-2025, which implies that the field of generative AI is rapidly becoming a mature and scholarly topic in academic libraries. The level of methodological difference is obvious: Lai (2023), Mugaanyi et al. (2024), and Wang et al. (2025) evaluate AI reference through experiments; large-scale quantitative surveys are represented by Choudhury and Shamszare (2023), Haris et al. (2025), and Elsayed et al. (2025); qualitative and interview-based Robertly by Chen (2025) and Kim (2025); and mixed-method designs are offered by combination These have been found on North America, Europe, Asia, Middle East and the Global South, which appears to make them have a global applicability rather than region-specific experimentation. Thematic areas of the research include (1) the quality and accuracy of references, (2) the roles and competencies required in the workforce, and (3) the trust of users and their adoption behaviour. It is worth noting that the volume of publications surged afterwards, after the widespread introduction of ChatGPT, thus creating an interdependent socio-technical evidence base, where experimental research identifies performance constraints, survey research identifies expectations and attitudes, and qualitative research clarifies how workforce mediation is generated under the threat of trust-sensitive reference risk.

Table 1. Characteristics table of selected Studies.