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

Following the screening stage, 30 reports were sought for retrieval. However, 9 reports were excluded because full texts were unavailable through open-access sources or institutional archives. Consequently, 21 full-text articles were assessed for eligibility. At the eligibility stage, 8 studies were excluded because they did not sufficiently address at least one of the core review dimensions, namely reference quality, librarian workforce roles, or user trust. Ultimately, 13 empirical studies satisfied all inclusion criteria and were included in the final systematic review (Table 2).

3.2. Quality Assessment and Risk of Bias

Overall, the quality appraisal indicates moderate-to-high methodological quality, with recurrent limitations in recruitment transparency, representativeness, and reliance on self-report. The cross-sectional and descriptive evidence comprised (Chigwada & Pasipamire, 2024; Deschênes & McMahon, 2024; Haris et al., 2025; Ismail et al., 2024; Elsayed & Abusharhah, 2025; Khan, 2025; Khan et al., 2026), and the survey findings summarised by (Grams, 2024). These studies were appraised with the CASP checklist (Appendix B). Recruitment strategies were frequently unclear or convenience-based. The survey summarised by (Grams, 2024) was limited by a small single-institution sample, while (Chigwada & Pasipamire, 2024) also had a small sample and limited analytical depth, increasing the risk of sampling bias.

The qualitative and documentary studies by (Chen, 2026; and Huang et al., 2023) were appraised with the CASP Qualitative Checklist. Both had clear aims and coherent designs. Chen’s interview study provided rich thematic findings but did not clearly report researcher-participant reflexivity. (Huang et al., 2023) used a transparent comparative sample of strategy documents from 50 universities, although limited reporting of coding reflexivity constrained assessment of interpretive influence. (Gmiterek & Kotuła, 2025) was the sole mixed-methods study and met most MMAT criteria, with limitations relating to institutional diversity and control of confounding factors. (Lai, 2023; and Mugaanyi et al., 2024) were appraised with the JBI quasi-experimental checklist. Both lacked a control group; pre/post measurement and follow-up were unclear for both; and Lai’s statistical analysis was rated only partially appropriate. These limitations support a moderate-to-high, rather than uniformly high, confidence rating and require cautious interpretation of generalisability 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. Across the included studies, the principal risks of bias included convenience and volunteer sampling, self-reported behavioural measures, limited representativeness, and the absence of longitudinal validation. Response bias and social desirability bias were highlighted in survey-based studies, but the absence of reflexivity reporting meant that there was moderate interpretive bias in qualitative studies. Experimental studies had similar evidence of not being subject to measurement bias, albeit specific to their contexts with limited external generalisability. No evidence of selective outcome reporting was found. Quality assessments for each study, as per their designated appraisal tool, along with an overall summary of quality assessments, are provided in Appendix 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 practice, and user trust. The study designs of the 13 major studies include experimental studies, large-scale surveys, mixed-methods studies, and qualitative case studies, thus demonstrating the complexity of using AI in libraries (Table 2). 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.

The included studies addressed diverse dimensions of generative AI integration in academic libraries, including reference accuracy, citation reliability, workforce transformation, governance preparedness, AI literacy, user trust, and adoption behaviour. Experimental studies primarily examined information integrity and reference quality, while survey-based and qualitative studies focused on professional readiness, governance concerns, and user perceptions across different institutional and regional contexts.

3.4. 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.

3.4.1. Scope and Distribution of the Evidence Base

The review is anchored on 13 peer-reviewed studies published from 2023 to 2026. Experimental or quasi-experimental studies by (Lai, 2023; and Mugaanyi et al., 2024) evaluated reference performance and citation reliability. Eight survey-based or descriptive studies. (Chigwada & Pasipamire, 2024; Deschênes & McMahon, 2024; Elsayed & Abusharhah, 2025; Grams, 2024; Haris et al., 2025; Ismail et al., 2024; Khan, 2025; and Khan et al., 2026) examined adoption, trust, ethical risk, AI literacy, and workforce readiness. (Chen, 2026; and Huang et al., 2023) provided qualitative interview and documentary evidence, respectively, while (Gmiterek & Kotuła, 2025) used mixed methods. The studies span Europe, Asia, the Middle East, Africa, and the Global South. Their thematic areas include reference quality, librarian roles and competencies, strategy and governance, and user trust and adoption behaviour. Together, the evidence base shows how performance constraints, user perceptions, institutional strategy, and workforce mediation interact.

3.4.2. Impact of Generative AI on Reference Quality and Information Integrity

In both the experimental and evaluative literature, a clear performance gradient is evident in generative AI-mediated reference services, and effectiveness depends heavily on question complexity. Experimental research using controlled testing with absolute reference questions shows that ChatGPT performs relatively well on low-complexity queries (directional, facilities-related, policy, and simple factual). In an experimental assessment of 58 real academic library reference queries, (Lai, 2023), the data indicate that accuracy was highest for low-READ-scale queries (Table 3). In contrast, in a qualitative study by (Chen, 2026), all librarians agreed that low-READ-scale queries could be effectively handled by AI, thereby reducing routine informational queries.

In comparison, high-complexity reference tasks cause a steep decline in performance. (Lai, 2023) notes that ChatGPT was rated lowest in higher research queries, known-item searches, and access to e-resources, particularly when a specific institutional context or disciplinary complexity was required. The survey evidence summarised by (Grams, 2024) is consistent with this pattern: 72.3% of students agreed that ChatGPT could not comprehend some questions and 66.6% agreed that its responses could be incorrect. The librarians interviewed in (Chen, 2026) also reported being reluctant to use AI in high-stakes reference interactions due to the risk of misinterpretation and partial contextual reasoning. The experimental methods show that retrieval accuracy decreases in proportion to the complexity of a query, and the qualitative data from practising librarians support this notion, indicating that these performance deficiencies are due to structural limitations and not temporary conditions. These quality limitations directly increase reliance on librarian mediation and shape users’ scepticism towards AI-mediated reference outputs.

Alongside the superficial quality of the answer, information integrity failures pose the most significant threat to AI generation in the academic reference domain. There is also quantitative experimental evidence of high percentages of citation hallucination and bibliographic inaccuracy. In the cross-disciplinary citation experiment by (Mugaanyi et al., 2024), only 32.7% of DOI references in natural sciences and 8.5% in humanities outputs were accurate, and hallucinated references in humanities-oriented outputs reached 89.4% (Table 4). These disciplinary differences show that generative AI is weakest in areas where citation conventions are heterogeneous and context-specific. These integrity failures help explain the relationship between citation inaccuracy, user scepticism, and governance concerns, linking reference quality directly to trust erosion and governance risk.

These results build upon previous issues raised in reference-service assessments. (Lai, 2023) also identified fabricated or misleading references in responses to research-oriented questions, but (Chen, 2026) states that librarians cited hallucinations as one of the primary reasons they should not use advanced AI in reference services. Combined, these studies demonstrate that plausibility rather than verifiability governs AI outputs another epistemic problem that the standards of academic libraries cannot accept. In short, experimental studies reveal that gains in speed and accessibility are achieved at the cost of epistemic reliability, a trade-off that survey-based optimism alone fails to capture. Despite the fact that generative AI helps to optimise the process and make it more user-friendly, it also threatens the principles of verification according to which scientific librarianship is established. (Deschênes & McMahon, 2024) disagree with the idea of unsupervised AI-based reference services, especially those managed by users who do not possess the necessary experience to identify the flaws. As a result, the necessity of the human validation layers can be found in the literature. Instead of doing away with reference services, (Chen, 2026; and Huang et al., 2023) found that AI remained weakly embedded in most formal academic-library strategies.

3.4.3. Transformation of Librarian Roles and Workforce Competencies

Generative AI redeploys librarian labour not as a proactive innovation strategy, but as a compensatory response to documented weaknesses in reference quality and information integrity. According to the interview-based results provided by (Chen, 2026), 100% of the reference librarian respondents (n = 25) stressed the importance of human validation when using ChatGPT, especially for complex or high-stakes queries, which would make librarians central filters of quality rather than dispassionate intermediaries. (Khan, 2025) found that GCC university librarians were already using GenAI for research queries, article summarisation, information retrieval, research writing, and service enhancement, but many depended on self-directed training because formal programmes and institutional support were uneven. (Khan et al., 2026) further showed that AI ethics competence and ethical governance significantly strengthened responsible-use intentions among 305 academic-library professionals. Similar findings were reported by (Ismail et al., 2024) in Pakistani academic libraries, where respondents expressed high levels of interest in AI integration (4.61/5) despite persistent infrastructural and training limitations, suggesting that institutional readiness remains lower than professional motivation. According to (Gmiterek & Kotuła, 2025), only 46.4% of Polish academic libraries today use generative AI but only 7.1% have official guidance on AI, and 77.8% cite the lack of employee competence as the most significant challenge. While survey studies report strong professional optimism, they simultaneously expose governance and skills gaps that limit safe operationalisation, revealing a disconnect between aspiration and institutional readiness. (Elsayed & Abusharhah, 2025) found that although 81% of respondents acknowledged potential ethical risks associated with AI, only 12% reported direct institutional experience managing AI-related ethical issues, suggesting a gap between ethical awareness and operational preparedness. The signs of misalignment can be seen at an early stage in the career ladder; according to (Chigwada & Pasipamire, 2024), over 70% of LIS students mentioned repeated misinformation, bias, and ethical uncertainty despite the heavy use of ChatGPT, which indicates that early adoption moves up the career ladder faster than formal teaching on AI literacy can do. These workforce pressures emerge directly from trust-sensitive reference risks, illustrating how declining epistemic reliability necessitates increased human mediation rather than automation (Fig. 3).