Developing university library services for supporting “AI for Science” research: Understanding the requirements of users

AI4S

AI is increasingly integrated into scientific discovery to enhance and accelerate research, assisting scientists in generating hypotheses, designing experiments, collecting and interpreting vast amounts of data and gaining insights unattainable through the traditional scientific method (Wang et al., 2023). AI4S now stands as the foremost scientific paradigm, signifying the profound convergence and deepening integration of AI with scientific research.

The very first definition of AI was proposed by McCarthy et al. (1956) at the Dartmouth Conference. In this proposal, AI was envisioned as “a machine” that was made to simulate and precisely describe “every aspect of learning or any other feature of intelligence.” Since then, a global surge in AI-related research has been initiated. The subsequent advancement of AI-driven scientific research has evolved through several sequential phases, and has now entered a profound Convergence Era (Mitchell, 2022; Russell and Norving, 2020).

This era is marked by the rapid and deep integration of advanced AI methodologies—particularly deep learning-into the discovery processes of nearly all scientific landscapes, accelerating the pace and broadening the scope of knowledge (Wang et al., 2023). As reflected by a growing body of recent studies, it is characterised not only by technological adoption but by a structural and multi-dimensional embedment of AI within scientific research activities.

First of all, AI-aided data collection and curation are considered constructive. Due to the substantial volume and complexity of the datasets generated from research activities, AI technologies underpin advanced data collection and curation, encompassing selection (Hinton and Salakhutdinov, 2006), annotation (Ratner et al., 2017), generation (Littman, 2015) and refinement processes (Bergenstråhle et al., 2022). Besides, AI creates conditions conducive to more meaningful representations of scientific data. Three emerging strategies, including geometric priors (Townshend et al., 2021), self-supervised learning (The LIGO Scientific Collaboration and the Virgo Collaboration, et al., 2017) and language modelling (Biswas et al., 2021), are utilised to retain as much information as possible while remaining simple and accessible.

Secondly, AI has demonstrated significant promise in testable hypotheses formulation, which remains a laborious process. Not only can AI methods generate hypotheses by identifying candidate symbolic expressions from uproarious observation, but also assist in research design and experimental evaluation (Wang et al., 2023). Furthermore, the advent of deep learning has positioned AI-driven experimentation and simulation as viable alternatives. Although experimental validation remains a critical yet costly, impractical, and potentially life-threatening endeavour (Dance, 2009), computer simulation, as a promising alternative, offers a promising approach to bridge observation with hypotheses, leading to more efficient testing and flexible experimentation (Gao et al., 2022; Segler et al., 2018).

AI technologies have fundamentally reshaped the research activities, introducing transformative paradigms. Accordingly, research support is expected to be provided from university libraries, which makes the transformation of RSS imperative and appealing.

RSS

RSS is defined as anything undertaken by a library to support and facilitate research within its parent institution (Hoffman, 2016; Voog and Wiklund, 2013). Over the past two decades, RSS have emerged as a significant research topic, reflected in a growing body of literature (Hussain and Rafiq, 2025). Since a well-organised support system is substantial for high-quality scientific research, university libraries are considered an indispensable component, playing a pivotal role in supporting scholarly activity and knowledge management (Chen and Zhou, 2021; Pulford et al., 2020).

Different from reference services, which typically do not involve librarians in the research processes, RSS positions librarians as members of the research team, contributing to various stages of the research lifecycle (Hoffman, 2016; Tang and Zhang, 2019). Rather than focusing primarily on information retrieval and discovery based on academic collections and resources, RSS encompass a range of supportive activities. Historically, university libraries have offered diverse forms of RSS, including subjective librarian services (Borrego et al., 2018), research data services (Tenopir et al., 2019; Zar Kyaw et al., 2025), digital scholarship services (Li et al., 2020; Zhou et al., 2019), bibliometrics services (Shoaib et al., 2023), open science services (Liu and Liu, 2023) and so on.

In addition to examining the content of RSS, existing literature also reveals empirical studies that investigated its implementation across different contexts. University libraries in countries such as the USA, UK, Australia, Canada, Switzerland, Netherlands, Singapore and China have made considerable efforts to offer RSS (Ho et al., 2026; Si et al., 2019; Tang and Zhang, 2019), especially those related to research data management (Tenopir et al., 2019; Zar Kyaw et al., 2025), open access (Liu and Liu, 2023), research consultations (Rogers and Carrier, 2017), scholarly publishing (Li et al., 2018; Lund et al., 2023), research tools (Siddaiah, 2017), and training (Hussain and Sudir, 2025; Taylor and Simon, 2026).

Existing literature suggests that RSS are challenging to investigate, identify and categorise due to their interconnected nature. Consequently, the research lifecycle model is frequently employed as a framework to examine RSS and their interactions, helping to develop a more comprehensive understanding of researchers’ requirements (Hussain and Rafiq, 2025; Si et al., 2019). The research lifecycle refers to a cyclical research process that consists of several primary phases, such as research idea proposal, research conducting and implementation, results publication and dissemination and preservation (Gessner et al., 2017).

To encompass all stages of the research process, several distinct research lifecycle models have been proposed in the literature. For example, Zhou et al. (2019) introduced a five-stage model that describes services corresponding to research ideas formulation, research collaboration construction, proposal, conduction and publication. Similarly, Maxwell (2016) divided research into phases of Planning, Project, Publication and e-Preservation, and outlines the services which the library should provide along with other potential partners on the campus.

Development of RSS towards AI4S

While research on RSS has been extensively discussed and is quite prevalent, existing studies have largely remained grounded in the data-intensive research paradigm (Zhou et al., 2019). The rise of AI4S has fundamentally reshaped research practices, from hypothesis generation and data analysis to scholarly communication, creating an urgent need for library RSS to evolve accordingly.

In response to this new paradigm, a growing body of literature has begun to discuss the changing role of libraries and the upgrading of library services in the context of AI4S (Cox, 2023; Hosier and Cantwell-Jurkovic, 2025; Jimenez et al., 2024), with the development of RSS recognised as a key component of this transformation (Hussain and Rafiq, 2025; Zhou et al., 2019). However, the majority of these discussions remain largely speculative. They tend to focus on generic prospects, potential scenarios and desirable future directions, primarily by theoretical reflections and speculation rather than grounded investigations of current practices (Cox et al., 2019; Hosier and Cantwell-Jurkovic, 2025; Razack et al., 2021).

What is notably absent in the literature is systematic, empirical research that moves beyond guessing and predicting, and engages directly with the realities of service implementation. In particular, there is a striking lack of in-depth, user-centred studies that explore emerging user requirements. What researchers truly need from AI4S-enabled RSS? How they interact with AI tools in their workflows? Whether they perceive libraries as relevant providers of such support? Without this empirical grounding, service development risks being disconnected from actual user practices.

Therefore, the concrete development, practical operationalisation, and sustainable maintenance of RSS in the AI4S era remain largely unclear and in need of further investigation, especially from a user-centric perspective.

Research design

To effectively respond to the research questions, this study adopted an inductive qualitative approach and included two inductive stages: a literature review and a case study. In the pre-review, a generic review was undertaken to understand RSS development in university libraries during the research paradigm shift. As a result, AI4S was considered to be a dominating research paradigm in the near future, which could significantly change scientific discoveries and research behaviours. It was indicated that user requirements should be comprehensively understood and captured, and meanwhile, long-term strategies of RSS towards AI4S have to be envisioned by university libraries.

The literature review was conducted during October and December 2024. Academic works in English and Chinese were retrieved, collected and analysed. Three academic databases were selected and systematically searched. Web of Science and Scopus were used and searched for the retrieval of English articles. Chinese works were collected from China National Knowledge Infrastructure (CNKI), the most popular and inclusive database of Chinese academic publications.

(“AI” or “AI for Science” or “AI4S” or “artificial intelligence” or “machine learning” or “data mining”) and (“research support” or “data management” or “digital scholarship” or “digital humanities”) and librar*

The search strategy was designed to be as inclusive as possible, with the purpose of including as many relevant and useful articles as possible. There was a chance that this inclusive strategy would return a substantial number of irrelevant publications. Thus, an exclusion criteria list was designed for manual screening: (1) duplicate articles from different databases would be identified and removed; (2) articles not specifically focused on RSS in university libraries would be excluded; (3) position papers and subjective opinion papers lacking adequate and justifiable theoretical support would also be discarded.

As illustrated in Table 1, the literature review provided an overview of general RDMS requirements that served as a basis for the case study investigation. In this part of the work, five themes of user requirements emerged.

OPEN IN VIEWER

Table 1.

An overview of RSS identified from the literature analysis.

Themes	Library RSS for AI4S Research	Supporting citations
Promoting AI4S readiness	Training and tutorial on AI4S tools	Çakmak and Eroğlu (2025), Michalak et al. (2025)
	Access to a full range of AI4S systems and tools	Michalak et al. (2025), Zhao (2024)
	Digital and physical spaces for AI4S research collaboration	Massis (2018), Osborne et al. (2024)
	AI4S literacy promotion	Chigwada (2024), Cox and Mazumdar (2024)
	Technological infrastructure and computing power for AI4S research	Chigwada (2024), Cox et al. (2019)
Formulating research idea	Literature and review support	Asemi et al. (2021), Gupta et al. (2022), Lappalainen and Narayanan (2023)
	Validation of research ideas/hypothesis	Garbuio and Lin (2021), Kulkanjanapiban et al. (2025)
	Bibliometric analysis of the research field	Cox et al. (2019), Yao et al. (2019)
	Grant seeking and proposal development	Hussain and Rafiq (2025), Singh et al. (2025)
Building collaborative network	Identifying potential research partners	Howlett et al. (2024), Huang (2018), Wu and Wang (2018)
	Team building and management	Dong and Ma (2019)
	Negotiation of roles and tasks	Kulkanjanapiban et al. (2025), Monyela and Tella (2024), Siemon (2022)
	Research progress coordination	Gupta et al. (2022), Jabeen et al. (2016)
Conducting research	Data collection, visualisation and analysis	Cox et al. (2019), Hussain and Rafiq (2025)
	Research data storage and management	Cox et al. (2019), Koltay (2019), Öztemiz and Şahin (2026)
	Ethics and compliance review	Çakmak and Eroğlu (2025), Mishra (2023)
	Digital project management	Cox (2023), Nenni et al. (2025), Vergara et al. (2025)
Publishing results	Academic editing and review support	Cox (2023), Hosier and Cantwell-Jurkovic (2025), Misra and Chandwar (2023), Razack et al. (2021)
	Journal recommendation and submission compliance	Kousha and Thelwall (2024), Ortega et al. (2018)
	Publishing and open access support	Jimenez et al. (2024)
	Research impact tracking and promotion	Gutzman et al. (2018), Ye (2019)

Interview data collection

The case study employed an inductive and exploitative qualitative approach, using a semi-structured interview and a thematic analysis. The semi-structured interview questions were specially designed according to the RSS requirements identified and listed in the literature review (summarised in Table 2). It was worth noting that the interview scripts were designed for two types of interviewees: one for researchers, the other for library management and librarians.

OPEN IN VIEWER

Table 2.

Interview questions.

Categories	Researchers	Library management and librarians
Generic question	Could you briefly describe your work situation? (e.g. name, age, research team affiliation, and current primary research focus?)	Please briefly describe your work situation (e.g. name, age, professional background, work department, position, years of experience, job responsibilities) and information about the library.
	How often do you utilise the university library’s RSS, and what types of services do you primarily use?	What new services have libraries expanded to offer in the context of AI4S? What additional AI-supported RSS scenarios remain to be explored by libraries?
Promoting AI4S readiness	Has the AI4S research paradigm had any specific impact on your research activities? (e.g. research efficiency, topic selection, or data processing)	How is AI transforming traditional RSS and users’ research? How should libraries respond to these changes brought by AI?
	What changes have you observed that AI has brought to RSS?	How has AI transformed training, tutorials or infrastructures? What new user needs have emerged under AI4S?
	Do the AI-related services offered by the library meet your needs? What other specific needs do you have?	How do you assess the prospects for integrating AI into services at this stage? (e.g. academic trend identification, AI-enhanced retrieval)
Formulating research ideas	How has AI transformed information integration, literature acquisition, and scientific novelty searches?	How has AI transformed information integration, literature acquisition or scientific novelty searches?
	Do you have new demands for RSS? What is the most pressing need you hope the library will address?	What new user needs have emerged under AI4S? What about the effectiveness and feedback?
	Have library services met your expectations in practice? If no, what caused it?	How do you assess the prospects for integrating AI into these services at this stage? (e.g. intelligent topic recommendations, literature analysis tools)
Building collaborative network	How has AI transformed the way you expand your network of collaborators and foster an environment of equal communication?	What changes has AI brought to expanding collaborations or building communication platforms?
	Do you have new demands for RSS? What is the most pressing need you hope the library will address?	What new demands have users expressed under AI4S? What is the most prominent demand? What about the effectiveness and feedback?
	Have library services met your expectations in practice? If no, what caused it?	How do you assess the prospects for integrating AI into services at this stage? (e.g. social network analysis, talent matching algorithms)
Conducting research	In the specific implementation of research projects, how has AI facilitated your work in areas such as data processing, writing guidance, and project management?	How has AI transformed data processing, research data management and project management?
	Do you have new demands for RSS? What is the most pressing need you hope the library will address?	What new user demands emerged under the AI4S framework? What about the effectiveness and feedback?
	Have library services met your expectations in practice? If no, what caused it?	How do you assess the prospects for integrating AI into services at this stage? (e.g. intelligent data processing, writing assistance and refinement)
Publishing results	When handling the publication of research findings, how does AI support you in areas such as publishing guidance, intellectual property management, and achievement evaluation?	What changes has AI brought to publication guidance, intellectual property management or achievement evaluation?
	Do you have new demands for RSS? What is the most pressing need you hope the library will address?	What is the most prominent demand? What about the effectiveness and feedback?
	Have library services met your expectations in practice? If no, what caused it?	How do you assess the prospects for integrating AI into services at this stage? (e.g. academic impact analysis, intelligent submission recommendations)
Open-ended questions	What risks and challenges do you believe libraries face when integrating AI into research support services? What are the reasons? (e.g. algorithm adaptability, data privacy, insufficient computing power, etc.)
	What’s your suggestion?

The interview was conducted in December 2024. The research team managed to reach 25 interviewees, including 18 researchers and 7 library management and librarians. The demographic information of the interviewees is shown in Table 3. Each interview lasted 22–70 minutes. With the participants’ consent, all the interviews were recorded and then manually transcribed into Word documents.

OPEN IN VIEWER

Table 3.

Demographic profile of participants.

Backgrounds	Number of participants
Professional position
Library management	2
Librarians	5
Researchers	18
Discipline
Library science	7
Law	6
Materials science and engineering	5
Accounting	2
Energy resources	1
Oceanophysics	3
Chemistry	1
Total	25

Interview data analysis

The analysis of interview data employed a thematic analysis method. Two thematic analysis tools were used: coding and theoretical representation. The analysis of the qualitative data used a qualitative data analysis application, MAXQDA 2022, to examine, interpret and constantly compare emerging codes. To include as many evocative codes as possible, the coding process was designed to be inductive. Inductive coding is exploratory and data-driven, allowing researchers to remain open-minded to new themes that may arise from the text data (Fereday and Muir-Cochrane, 2006).

To ensure methodological rigour in textual data coding, this study employed an approach encompassing three established strategies. (1) Theoretical foundation. As mentioned in the preceding section, the coding is based on a prior development of theoretical propositions to guide data collection and data analysis (Yin, 2003). (2) Iterative coding development. After pre-coding of several interview scripts, all preliminary codes derived from initial transcripts underwent three revision cycles—provisional, revised and final. (3) Integrated coding methods. The interview transcripts were coded using three types of coding: open coding, axial coding and selective coding. As illustrated in Table 4, the analysis of interview data pointed to 10 user requirements of RSS, which emerged in five main themes: promoting AI readiness, formulating research ideas, building collaborative networks, conducting research, and publishing results.

OPEN IN VIEWER

Table 4.

RSS requirements and main themes.

Themes	User requirements	Supporting quotations
Promoting AI4S readiness	AI education	• Training and tutorial on AI4S tools • AI literacy promotion
	Infrastructure and computing power	• Specialised models • Stable access • Integrated platforms
	Resource construction	• Digital databases and resources • Personalised information push • AI-translated resources
Formulating research ideas	Literature and review support	• Information retrieval • Information screening
	Validation of research ideas/hypothesis	• Bibliometric analysis • Sci-tech novelty retrieval • Validation of the hypothesis
Building collaborative network	Identifying potential research partners	• Platform • information
Conducting research	Data processing and management	• Data visualisation • Data analysis • Data storage • Data sharing
Publishing results	Review support	• English proofreading • Text editing
	Submission compliance	• Journal recommendation • Result evaluation
	Property protection	/

Promoting AI4S readiness

Promoting AI4S readiness is considered “necessary” (Interview 10 and 19), “extremely significant” (Interview 5), “effective” (Interview 18 and 19), and useful (Interview 20 and 21). The interview data analysis revealed three RSS requirements:

It is arguably that AI-supported search does not remove the need for information literacy. Since data and information provided by AI can be “inaccurate” (Interview 19), AI4S literacy is relied on in the process of using AI as an auxiliary tool to enhance research efficiency (Interview 18). It should be highly noted that the capability of using AI correctly determines how effectively it can contribute to scientific research. In this case, XUL are born to serve as an advisor and educator. As shown in the interview data, there is a growing preference for “lectures with concrete themes” (Interview 14 and 21), including “guidance on writing formal and standard text” (Interview 9) and “training on retrieval tools and databases” (Interview 11, 18, and 20), rather than comprehensive, systematic training programmes (Interview 21). While a wealth of video tutorials is provided on online platforms, live lectures are considered more beneficial, as they are designed for more “personalised interaction and immediate feedback” (Interview 21).

Development of resource construction, especially digital academic resources and personalised information recommendations, is requested by most participants. “Digitising books is essential, as it serves a foundation for supplementary reference materials . . . By employing AI technologies, standardised keywords can be indexed, and full-text retrieval should be enabled, thereby achieving deep utilisation of library resources” (Interview 14). Drawing on a broad collection of user data of library resources, AI are able to provide personalised recommendations for resources (Interview 4, 16, 19, 21) and news (Interview 12) that tend to align with researchers’ research interests and individual preferences.

But I think the best aspect of AI is its ability to predict—it can anticipate what you might need in the future. I believe this is where it truly excels. Take TikTok, for instance. The precise recommendation is a result of the AI algorithms that can tailor content to users’ preferences. This is why I see it as a way that could significantly enhance researchers’ experience (of library services) in the future. (Interview 16)

Although technological infrastructure and computing power are the foundation of AI4S RSS, stable access to AI4S systems and tools can be challenging, which makes researchers think about the possibility of “purchasing access to commercial general language models” (Interview 10). Especially, integrated AI models into the library platform, functioning as an AI assistant for databases, are mentioned. This service is designed to “accelerate the learning curve for research novices, enabling them to more rapidly navigate and utilise database resources effectively” (Interview 11). Furthermore, interview participants demonstrated a growing demand for domain-specific models, which would outweigh their reliance on general large language models to some extent. Trained on dedicated literature from various disciplines, these specialised models would be capable of delivering “more expert and precise information” (Interview 16) within their respective domains.

To our expectation, whether specialised models can be developed is interesting. For instance, services tailored for computer science, automation, or electrical engineering would be quite distinct. If it can offer such differentiated services and resources, I believe they would be quite useful. (Interview 16)

Formulating research ideas

As the start stage of the research process, research idea formulation based on numerous and various resources can be consequential and affect the following research processes. The interview data analysis revealed two major RSS requirements:

Literature and review support services are expected, including AI retrieval and AI screening. Allowing natural language search and multiple rounds of dialogue, AI retrieval is deemed “a smarter search engine” (Interview 8). The most paramount benefit lies in search convenience enhancement, “developing a systematic search strategy” (Interview 15) and “enabling precise results with just a rough description” (Interview 18), which can save a substantial amount of time. Though initial search results tend not to be highly accurate, with multiple rounds of questioning and refinement, it is able to provide a relatively useful answer ultimately. Furthermore, the main ideas of literature are captured and summarised by AI to “facilitate efficient screening” (Interview 14, 21) and provide “a barely satisfactory literature review” (Interview 5, 17)

I also use AI to ask questions during my regular studies. Traditionally, we’d use search engines to look things up, but often the results we get aren’t very relevant to the answers we’re looking for. However, when using AI now, the content it provides is usually quite relevant. And even if it’s not, you can tell it how to adjust, and it generally gives you a pretty satisfactory response. In this sense, it largely replaces search engines by consolidating information for you. (Interview 21)

Furthermore, AI-based RSS can play a role as a collaborative partner for brainstorming (Interview 20). As revealed in interviews, researchers expect to engage with AI to discuss and validate research ideas, including “the selection of analytical theories” (Interview 14). Since the imperative to establish a project’s innovation, significance and practical feasibility presents a major challenge, AI tools offer significant support by helping researchers to critique and validate their hypotheses, thereby refining research questions and enhancing the overall robustness of study plans from their inception.

The predictive capacity of AI constitutes its fundamental advantage. By analysing historical data to evaluate the potential of ongoing research and estimate its success probability, AI directly addresses the inherent uncertainties in innovation. This capability for probabilistic reasoning distinguishes it from mere classification algorithms. Therefore, leveraging AI for forecasting project outcomes prior to their initiation—and even prior to human ideation—represents a promising direction, as it can guide resource allocation and prevent redundant efforts. (Interview 16)

Building a collaborative network

Conventionally, research collaboration relies on ingrained cognitive patterns and pre-existing rational networks. Yet, the growing emphasis on interdisciplinary studies necessitates cross-disciplinary collaborations to explore unfamiliar fields, generating a pressing need for comprehensive intelligence on potential partners and their research domains. The requirement of identifying potential research partners emerges from the interviews.

According to interview participants, XUL is expected to develop a scholarly collaboration platform, which will display the research achievements and interests of prominent scholars across disciplines, achieving a consolidation of academic profiles (Interview 10, 19). By leveraging extensive data collection, analysis and matching algorithms, the platform is designed to “perform exact queries and personalised recommendations” (Interview 10), effectively mitigating the transaction costs associated with identifying suitable research collaborators.

However, despite the algorithmic recommendations, a notable portion of respondents expressed neutrality or scepticism. While AI matching relies on quantifiable criteria, successful collaboration is inherently limited by these algorithms and is ultimately contingent upon in-depth and interpersonal communication (Interview 16).

Conducting research

widely regarded as the central stage of the research process. This centrality necessitates robust research support services, primarily encompassing data management and ethics and compliance review. Libraries are requested to provide research data management services to support data-centric research. Collected research data needs to be processed, analysed and visualised through a set of procedures with the assistance of librarians. The initial phase of data processing, particularly the manual extraction of information from large volumes of unstructured text and massive amounts of data, tends to “involve tedious, repetitive and time-consuming tasks” (Interview 21). AI-driven approaches, especially those leveraging large language models, offer a powerful solution as the task aligns closely with their core functional strengths in pattern recognition and information retrieval (Interview 13, 14).

During data analysis, executable code frameworks can be generated by AI from natural language descriptions (Interview 14). Although subsequent modifications are often necessary, this capability significantly accelerates the development. Particularly in complex analytical scenarios, AI can perform analyses that might be beyond a researcher’s technical skills (Interview 15), thereby empowering a broader range of scholars’ engagement in complex research.

Data processing services are likely needed across the humanities, social sciences, and STEM fields today. For quantitative analysis, I believe AI tools can genuinely assist you in writing code, which I find quite helpful. Personally speaking, I think that having no understanding of languages like Python is still problematic. You need to know some fundamentals, then combine that knowledge with these tools to complete your analysis. (Interview 14)

AI demonstrates a significant capacity for synthesising professional technical roadmaps from user-described concepts. While human input remains essential for defining the conceptual direction, AI excels at translating these ideas into visually coherent and aesthetically structured diagrams, overcoming the challenges of unclear manual sketches. The primary advantage of this approach lies in “enhancing the comprehensibility and communicative efficacy of complex technical information” (Interview 21).

Publishing results

At the stage of concluding research, several requirements of RSS emerge before results publishing, including:

For non-native researchers, “academic writing in English poses a considerable challenge” (Interview 11), which may potentially fail to reflect the intended meaning and research findings. Before the application of AI tools, researchers may be compelled to resort to professional editing agencies, thereby creating a dual burden of significant time commitment and financial cost (Interview 14). Consequently, this concern drives a massive demand for libraries to provide AI-powered writing assistance tools, which can help scholars align their papers with general academic writing standards (Interview 7, 16, 19, 21).

AI excels at polishing academic papers. For instance, upon completing a project, AI can help optimise the content and structure of your final report. Overall, I find AI to be a highly useful tool—the key lies in how effectively you utilise it. (Interview 19)

The identification of a suitable publication journal is a critical and time-consuming step in the research lifecycle. As revealed in the interviews, researchers expect RSS to evolve from merely providing basic journal information to recommending aligned journals, especially English journals, based on manuscript titles, abstracts and keywords (Interview 16, 19). Journal information curated by libraries is generally perceived as highly authoritative and accurate. Thus, developing a systematic framework for journal recommendation by “establishing specific criteria to various departments and disciplines” (Interview 20) could provide significant value.

Furthermore, although concerns exist regarding data security when submitting sensitive research details to AI systems, there is a “greater willingness to trust AI applications that are locally deployed and institutionally supported by libraries” (Interview 15). Such assistance in publishing results is designed to save researchers significant time, thereby enhancing both efficiency and the potential for acceptance.

The protection of intellectual property rights enters a critical yet challenging phase following research publication. To address this, AI-based monitoring systems offer a promising approach to address this gap by enabling automated scanning for potential infringement across large-scale digital sources, thereby supporting rights enforcement (Interview 5). While the final determination regarding infringement and subsequent enforcement actions appropriately remains with researchers, RSS are expected to perform preliminary assessments of suspected cases.

Challenges of RSS development

Firstly, it is noteworthy that although the transformation and development of RSS in university libraries are imperative to address evolving user requirements in the AI era, the accompanying difficulties and challenges remain formidable and cannot be overlooked. These challenges now constitute significant impediments to the library’s institutional transformation.

The technical limitations of AI algorithms are evident and cannot be overlooked, leading to user concerns and scepticism about AI4S. Highly dependent on their training datasets, AI tools exhibit different performance. Particularly in scientific research, without reliable academic databases as crucial learning material, their outputs tend to be “inaccurate” (Interview 9, 16, 22), “incredible” (Interview 12, 15, 16, 17, 18) and “obsolete” (Interview 7). Furthermore, as revealed by interview participants, data leakage is inevitable. Despite providing explicit directives against utilising conversation content for model training, users “remain unaware whether such data is ultimately exploited” (Interview 14). Also, no technology company can offer absolute assurance against data leakage, given the inherent bugs present in any complex technological infrastructure.

Research ethics faces significant and novel challenges with the integration of AI tools. Taking literature reviews as an example, “the direct incorporation of AI-generated literature reviews raises questions about plagiarism” (Interview 10). Besides, whether the formal acknowledgement and disclosure of the use of AI to assist manuscript editing is mandated and obligated is still pending. Although disclosure requirements for AI assistance are advocated by some scholars, “the operational implications and associated costs require further systematic investigation” (Interview 16). Particularly, reliable verification of such usage remains challenging.

Another significant concern is whether the use of AI truly benefits research fostering. As evidenced by the interview data, a notable contingent of scholars—particularly those engaged in educational roles—expressed apprehension that excessive dependence on AI-driven tools may inadvertently lead to the erosion of autonomous critical thinking and analytical reasoning capabilities. This apprehension underscores a deeper anxiety regarding the potential trade-off between technological convenience and the preservation of essential cognitive skills that form the cornerstone of rigorous academic inquiry.

Pursuit of personalised services

The advent of AI technologies has intensified the user requirement for specialised information and services. While significant requirements for RSS are acknowledged, concerns regarding its effectiveness and quality remain. A prevalent view among researchers is that “existing library services are perceived as largely rudimentary” (Interview 15, 17), mainly beneficial for research novices or users with limited AI proficiency. RSS addressing highly specialised or complex research inquiries is either underdeveloped or, where available, exhibits constrained effectiveness.

Consequently, users are demonstrating increased demand for more personalised RSS. They seek resource recommendations with greater domain-specific relevance (Interview 21). Furthermore, compared to general-AI large language models, researchers show “stronger preference for specialised models” (Interview 5, 8,14), which are “trained on authoritative academic datasets” (Interview 16) that offer enhanced accessibility within the research lifecycle. Indeed, the development of domain-adapted AI tools is perceived as better aligned with researchers’ operational requirements.

Theoretical contributions

The findings presented above not only summarise user-derived requirements for RSS but also offer several points of critical departure from prior research, thereby advancing our understanding of library service transformation in the AI4S era. While existing literature has extensively discussed the potential of AI to reshape library services (Cox, 2023), such discussions have largely remained speculative. In contrast, our findings provide empirical evidence that this potential is significantly constrained by tangible challenges, including technical limitations, ethical concerns, and fears of cognitive erosion, which may fundamentally affect user acceptance of AI-enabled RSS. Furthermore, our study challenges the assumption that researchers will naturally turn to libraries for AI-driven support. As shown in previous discussion, researchers perceive existing library services as rudimentary and prefer domain-specialised models trained on authoritative academic datasets, suggesting that libraries cannot rely on their traditional legitimacy but must actively reestablish relevance in the AI4S era.

Our study also contributes to the emerging discourse on AI ethics in libraries by grounding ethical concerns in actual user voices. While issues such as data leakage, plagiarism and disclosure requirements have been recognised in previous work (Van Noorden and Perkel, 2023), our findings reveal that these concerns directly shape researchers’ willingness to engage with library-mediated AI services. Additionally, the apprehension about the erosion of individual critical thinking skills, particularly among educator-researchers, adds a new dimension largely overlooked in library and information science literature. Collectively, these contributions reframe the research problem: the question is not simply how libraries can develop AI-enabled RSS, but whether and under what conditions users will actually adopt such services.