AI for research support

The distinction that matters most for research support is between AI tasks grounded in content the patron has already retrieved and AI tasks that require generating facts from training data. In research support, AI delivers genuine value mostly in the first group: helping patrons think through a research problem, processing documents they already hold, and explaining concepts they have not yet encountered. Understanding this distinction lets librarians steer patrons toward productive AI use without endorsing the problematic uses that lead to fabricated citations.

Search strategy development is the area of most consistent genuine value. When a patron brings a research question and needs a search strategy, AI can rapidly generate a substantive set of search terms, synonyms, related concepts, Boolean combinations, and relevant subject headings. For example, a librarian working with a graduate student on telehealth adoption and rural health equity can ask Claude to generate search terms organized by concept cluster in approximately thirty seconds. The clusters might cover telehealth terminology, rural health identifiers, equity and access language, and relevant MeSH headings for PubMed. This brainstorm does not replace the librarian's judgment about which terms suit which databases. It provides the raw material for that judgment with much less effort than generating it manually.

Summarizing provided documents is the second reliably useful application. A patron may have a set of abstracts, a full article, or a selection of PDFs and need help identifying key themes, methodological approaches, or relevant findings. AI working with that provided content produces reliable and useful summaries. This is the information-in mode discussed in Module 02: the model processes text the patron has already retrieved from authoritative sources rather than generating facts from training data. For example, a nursing student preparing a systematic review can paste twelve abstracts into Claude and ask it to identify shared themes, divergent findings, and methodological differences across the set. The structured synthesis that comes back would take considerably longer to produce manually. These summaries still warrant review rather than uncritical acceptance, but the hallucination risk is substantially lower than when AI generates content from scratch.

Explaining research concepts on demand is a third consistent strength. AI can explain research methods, statistical approaches, citation conventions, and disciplinary norms at the right level when a patron first encounters them, without a librarian being available. For example, a social work student unsure how a systematic review differs from a scoping review can get a clear, example-grounded explanation in plain language immediately. This on-demand explanation is particularly valuable outside library hours or when a patron is working independently between consultations. The important caveat is that any substantive claims in the explanation should be confirmed against authoritative methodology documentation before being relied on.

AI also serves research support well at the question formation stage. A patron struggling to articulate a focused research question can ask AI for five alternative framings of a broad topic, or for aspects of the topic that appear underresearched. The model's knowledge of any field has limits, including a training cutoff date and representation biases toward English-language and Western sources. It can still reliably generate the conceptual variety that helps patrons move from a vague interest to a workable question. This help is often invisible as "research support." Yet it meaningfully expands the librarian's consultative role when AI is positioned as a thinking partner in the consultation rather than a source of authoritative answers.

The failure modes of AI in research support are structural, not incidental. They follow directly from how large language models work, by prediction from training data, and they do not disappear with newer model versions or more careful prompting. To advise patrons effectively, librarians need to understand these failure modes well enough to explain the underlying reason, not merely list prohibitions. Patrons who understand why AI fails in specific ways develop more durable verification habits than patrons who have simply been warned.

Finding sources is the most consequential failure mode. A general-purpose AI chatbot cannot search databases. When asked to find peer-reviewed articles on a topic, the model generates credible-looking citations from statistical patterns in its training data. Some of these citations correspond to real publications; many do not. The fabrications are not random. They closely follow the conventions of real citations, including realistic author names, appropriate journal titles, and plausible publication years, which makes them hard to spot without verification. For example, a history student who asks ChatGPT for primary sources on the 1930s labor movement may receive a list mixing real archives, fabricated archive names, and citation details that are internally plausible but externally unverifiable. The danger is not that the citations look wrong; it is that they look right.

Current literature presents a related problem. Most major AI tools have a training cutoff date, after which no new information was incorporated, and it typically lags the present by one to several years. For patrons researching fast-moving topics such as clinical trials, pharmaceutical approvals, policy changes, or recent legislation, AI is actively misleading. The model describes, with full confidence, a landscape that may have changed substantially. For example, a policy student researching current federal AI regulation who asks Claude for an overview will receive information current at the model's training cutoff. Nothing in the response signals how much the regulatory environment has shifted since. Outdated information of this kind is particularly dangerous because it arrives in the same confident, well-organized prose as accurate information.

Specialized and niche subjects are underrepresented in AI training data in proportion to how little was written about them during the training period. This affects research support unevenly. A librarian helping with a mainstream topic in psychology or history can expect more complete coverage than one supporting research on an emergent subfield, non-English-language scholarship, or a topic whose primary literature lives in specialized professional databases rather than on the general web. For example, a patron researching Indigenous language preservation in a specific region may find that AI synthesizes mainly from a few English-language sources. The response then reflects an anglophone framing of the subject rather than the community-centered scholarship that would be most appropriate. Librarians should treat AI coverage of any specialized or underrepresented topic as presumptively incomplete.

Evaluating source quality is a task AI cannot perform reliably. A patron who asks AI to pick the most credible or methodologically rigorous sources from a list is asking the model for a judgment it has no mechanism to exercise. A language model evaluates sources the way it evaluates everything, by matching patterns to what sounds authoritative. It may rate a source as credible because of prestige vocabulary in an abstract, not because of any actual assessment of methodology, peer review status, or scholarly standing. To protect patrons from this misunderstanding, make the instruction point explicit: AI can describe what a source claims; it cannot assess whether those claims are warranted. Source evaluation remains a library professional competency, however sophisticated AI responses become.

Search strategy development is where AI most clearly earns its place in the consultation workflow. This section covers precisely how it contributes and precisely where it stops. The key distinction, worth repeating to patrons and colleagues alike, is that AI helps with thinking about how to search, not with searching itself. It generates terms, structures, and conceptual frameworks. It does not retrieve records, query databases, or access controlled vocabularies in real time.

The core workflow proceeds in stages that the librarian can model explicitly in a research consultation. First, articulate the research question clearly; the more specific the patron's question, the more useful the AI-generated search structure will be. For example, a question stated as "the effects of nurse to patient ratios on patient outcomes in intensive care settings" will produce more targeted search terms than "nursing and hospitals." Second, ask AI to generate search terms organized by concept cluster: "Generate search terms for each concept in this research question, including synonyms, related terms, alternate spellings, and relevant controlled vocabulary for PubMed." This prompt typically produces a structured set of term clusters that the librarian and patron can review and refine together.

Third, ask AI to suggest Boolean combinations and search string structures. For example, prompting Claude with "Suggest three Boolean search strings using these term clusters, with AND connecting the concepts and OR connecting synonyms within each cluster" produces ready-to-test strings. The librarian then evaluates these strings against the specific database's indexing conventions (MeSH terms for PubMed, Thesaurus descriptors for CINAHL, controlled vocabulary for PsycINFO) and adjusts them. AI can suggest that a medical search should include MeSH terms. It cannot guarantee that the specific terms it names are current and correctly formatted. That verification is the librarian's contribution, not an optional extra.

Fourth, iterate in response to database results. If an initial search returns too many or too few results, or lands in the wrong subarea of the topic, the patron can describe this to AI and ask for revised strategies: "My search for these terms returned 3,000 results in CINAHL, most of which are not focused on the intensive care setting. Suggest three strategies for narrowing: adding limiters, adding additional concept terms, or restructuring the Boolean." This iteration makes the consultation more dynamic and often surfaces search angles neither the librarian nor the patron had considered. Think of AI here as a well-read colleague who has absorbed an enormous amount of methodology literature and library instruction content. That colleague can generate options quickly, but the librarian's knowledge of specific database indexing is what makes those options actionable.

AI can also help patrons develop search strategies for databases they have never used. It can explain how a database is structured, what types of sources it covers, and how its search interface differs from familiar tools. For example, a social science student used to Google Scholar who is meeting Sociological Abstracts for the first time can ask AI to explain the database's scope, subject heading structure, and most effective search approaches for the discipline. This orientation shortens the learning curve for new database users in ways that complement rather than replace library instruction. It positions AI as a preparation tool for library work, not a replacement for it.

The information-in mode of AI use, providing documents for AI to process rather than asking it to generate information from training data, is the highest-reliability application in research support. It is worth developing as a core consultation tool rather than treating it as an incidental capability. When a patron brings abstracts, full articles, PDFs, or reports to an AI interaction and asks for help processing them, the hallucination risk drops substantially, because the model is working with provided text. This should shape how librarians introduce AI to patrons at the research support stage: emphasize working with sources the patron has already verified over asking AI to generate sources from scratch.

The most immediate application is abstract triage. A patron who has run a database search and retrieved thirty abstracts faces a time-intensive reading task. They must work through all thirty to decide which articles are worth pursuing to full text, which are tangentially related, and which are not relevant at all. AI can make an initial triage pass, organizing the full set of abstracts by relevance, methodology, or research focus. For example, a graduate student can paste thirty PubMed abstracts into Claude and ask: "Organize these abstracts into three groups: highly relevant to a systematic review on this topic, moderately relevant and worth reviewing in full text, and tangentially related. For the highly relevant group, note what each one contributes." This triage turns a forty-five-minute reading task into a ten-minute review, freeing the patron's time for deeper engagement with the most relevant sources.

For more advanced synthesis tasks, AI working with provided documents can produce structured syntheses that support rather than replace the patron's own intellectual engagement. Such tasks include identifying shared themes across a literature, mapping methodological approaches, or characterizing the evidence base for a specific intervention. To use this capability responsibly, librarians should explain to patrons that AI synthesis of provided documents is a starting point for their own analysis, not a finished product. For example, a public health student asking AI to synthesize the methodology sections of twelve provided articles may receive a useful overview of study designs, sample sizes, and outcome measures. The student must still engage with that synthesis critically and confirm key details against the original documents before using it in a literature review.

The limit of this approach is important to communicate explicitly. AI processes provided text and reflects it back in organized form. It has no access to the broader literature and cannot tell the patron what the provided documents omit. A patron who synthesizes fifteen articles and concludes that the literature "shows X" has synthesized fifteen articles, not the field. For example, if all fifteen articles happen to represent one methodological tradition or one disciplinary perspective, the synthesis will reflect that perspective without noting its limits. AI's facility with provided documents does not displace the librarian's role here: pointing out that a broader literature search should precede or accompany AI synthesis, and that the patron's selection criteria shape what the synthesis represents. That role becomes more important, because the ease of AI synthesis can create an illusion of comprehensiveness that a narrowly selected document set does not warrant.

General-purpose AI tools such as ChatGPT, Claude, and Gemini were designed for broad conversational use and carry the research risks described in this module's earlier sections. A distinct category of tools has been built specifically to address those risks by grounding AI responses in actual database searches or controlled scholarly corpora. For certain research support tasks, these tools are a meaningful improvement over general-purpose AI. Librarians who understand their capabilities and limitations can recommend them credibly to patrons in contexts where fabricated citations would be professionally or academically costly.

Connected Papers is an AI-enhanced discovery tool that visualizes the citation network surrounding a specific paper. A patron enters a known relevant article, and the tool maps both the papers that cite it and the papers it cites. The resulting visual graph makes the intellectual neighborhood of a topic navigable. For example, a sociology graduate student who has identified one foundational paper can use Connected Papers to surface the scholarship that has engaged with it. The student can discover articles that keyword searching missed and map the structure of the scholarly conversation in the field. This kind of citation network visualization is particularly valuable early in a literature review, when a patron needs to understand the landscape before committing to a specific angle. Connected Papers does not generate text from training data; it retrieves from real citation networks, so it carries no hallucination risk for bibliographic purposes. It shows connections among documents that actually exist.

Elicit is an AI research assistant that searches Semantic Scholar, a database of over 200 million academic papers. It returns real papers with AI-generated summaries drawn from the actual document content. For example, a clinical librarian helping a patron with a systematic review can use Elicit to run structured searches, organize results by study design, and generate comparative summaries across a retrieved set of papers. The difference from general-purpose AI is fundamental. Elicit grounds its responses in real search results from a controlled scholarly database, which substantially reduces hallucination risk for bibliographic content. This grounding does not eliminate all risk. AI summaries of real papers can still mischaracterize nuanced findings, and Semantic Scholar's coverage is not equivalent to the major licensed research databases. It is still a far more reliable starting point for literature review support than asking ChatGPT to find sources. Elicit is most useful for systematic review workflows, rapid literature mapping, and extracting specific data points across a large set of papers.

Consensus takes a complementary approach: it searches peer-reviewed papers and attempts to synthesize the overall state of evidence on empirical questions. For example, a patron asking "Does sleep deprivation affect academic performance?" receives a structured response. It shows how many papers Consensus found, how the evidence divides across positive and null findings, and links to the actual papers supporting each position. This evidence mapping is most useful for quick evidence checks, when the patron needs a rapid orientation to the research before pursuing a deeper literature review. The free tier is genuinely usable for most research support purposes, and the tool handles empirical questions in medicine, psychology, and social science well. It is less effective for humanities topics, historical questions, and highly specialized or emergent research areas that are poorly represented in its corpus.

Perplexity in academic mode combines AI synthesis with real-time web search and includes citations for every response. For research support, it is useful for factual orientation: helping a patron understand the policy context for a research topic, identifying key organizations in a field, or getting a quick overview of a methodology. The citations let the patron verify the AI's characterizations against primary sources. That makes this verification-enabled use more appropriate than general AI use for the orientation stages of research. Perplexity does not, however, replace database searching for peer-reviewed literature, and its summaries carry the same risk of mischaracterization as other AI-generated content.

Before recommending any of these tools to patrons, librarians should try them personally on topics they know well. That way they can evaluate the accuracy of summaries and the quality of retrieved results before directing patrons to rely on them. For example, checking whether Elicit's summaries of articles the librarian has already read are accurate produces the calibrated, firsthand judgment that credible patron recommendations require.

A second category of AI-native tool operates not over an external corpus like Semantic Scholar but over the databases and discovery layers the library already provides. This category has moved into production over the past year. Ex Libris's Primo Research Assistant performs retrieval-augmented generation over the Central Discovery Index and returns a small set of the most relevant sources, with citations back to records the patron can open directly. Clarivate offers a parallel Summon Research Assistant for institutions on that discovery platform. JSTOR, EBSCO (as AI Insights), Ebook Central, and Scopus have each added AI assistants that operate over their own indexed content. Their summaries and suggested sources are traceable to records the library's subscriptions actually cover. For example, a patron in EBSCO who asks AI Insights to summarize an article's argument receives a summary grounded in that specific article, not in a language model's general training data. That is precisely the distinction this module has drawn between AI working with provided or retrieved content and AI generating from training data alone. Before recommending these tools, librarians should check which are active under their institution's subscriptions, since rollout has varied by vendor and license tier. Where available, they belong alongside Elicit, Connected Papers, and Consensus. They have the added advantage of operating inside resources the library already licenses and the patron is already authenticated into.

A further development worth knowing about, even where it is not yet in routine use, is Elsevier's Deep Research feature within Scopus. It is an agentic research tool grounded in the Scopus database. Rather than returning a single response, it plans and executes a multistep research process and displays its reasoning on screen as it works: which searches it is running, which sources it is consulting, and how it is narrowing toward a synthesis. Module 01 introduced the vocabulary for this category (agentic AI: systems that set goals, plan tasks, and act with minimal guidance), and Deep Research is a concrete instance of it in research support. For a patron, watching an agentic tool's reasoning unfold is a different experience than receiving a finished answer. Each step, each search run, and each source consulted is visible and could in principle be checked. That visibility does not mean the output needs less verification. A multistep process grounded in Scopus still ends in a synthesis the patron must evaluate with the same judgment this module has built throughout. What changes is the texture of the interaction, not the professional standard applied to it. However autonomous these tools become, someone must still decide whether a synthesis actually answers the patron's question, represents the literature fairly, and is ready to inform their work. That judgment belongs to the patron and the librarian supporting them, not to the tool that produced it.

The most effective way to communicate AI's role in research to patrons is not a list of rules but a map of the research workflow, marking AI's appropriate and inappropriate uses at each stage. A map makes the distinction concrete and contextual rather than abstract. It can become a shareable handout, a consultation framework, or an instructional tool, depending on the librarian's setting and patron population.

A research workflow for most undergraduate and graduate work moves through recognizable stages: topic development and question formation, literature search and retrieval, source evaluation, reading and synthesis, and writing and citation. AI's role differs at each stage. Once the stages are mapped explicitly, the distinctions are not hard to communicate.

At the topic development stage, AI is genuinely useful. Exploring a topic's scope, generating potential research angles, identifying related fields and interdisciplinary connections, and narrowing a broad interest to a focused research question are all tasks where AI's brainstorming capabilities serve the patron without requiring verified factual claims. For example, a first-year student with a broad interest in climate change can use AI to narrow toward a specific angle (the intersection of climate change and food security in sub-Saharan agriculture). The student then brings that focused question to the library for database searching. This use is entirely appropriate and a net gain in research preparation.

At the literature search and retrieval stage, AI's appropriate role is limited to search strategy development: generating term clusters, Boolean combinations, and database recommendations. It does not extend to finding sources. The student takes the search strategy to actual databases: JSTOR, PubMed, PsycINFO, or whatever the appropriate disciplinary resource is. The instruction for this stage can be stated simply: "Use AI to develop your search terms; use the library database to find the sources." This clear two-step instruction prevents the most common and consequential research workflow error.

At the source evaluation stage, AI should not be used for quality assessment, for the reasons described earlier in this module. The patron must evaluate sources using conventional information literacy criteria (authority, accuracy, currency, coverage, and purpose) applied to the source itself. That evaluation requires direct engagement with the source. It cannot be outsourced to a model that cannot access the actual document or assess its scholarly standing.

At the reading and synthesis stage, AI is most useful once the patron has retrieved and verified the sources. Summarizing an abstract, explaining a methodology section in plain language, identifying the key argument of a dense theoretical paper, or synthesizing themes across a provided set of abstracts are all appropriate uses that work with documents the patron already holds. For example, a student who has retrieved twelve legitimate articles and needs to organize them for a literature review can use AI to produce an initial thematic map from the abstracts. The student then reviews, adjusts, and develops that map through their own analysis.

At the writing and citation stage, the primary guidance is twofold. AI-generated citations must never appear in final work without independent verification, and AI assistance with prose should be disclosed according to the institution's and instructor's policies. To prevent citation fabrication at this stage, treat verification as a required step rather than an optional precaution: before any citation goes into a paper, the patron must locate the actual document in a library database. Turning this map into a printable or digital handout for research consultations is among the most immediately practical steps a librarian can take to improve research AI literacy. A workflow framework gives patrons something concrete to follow, rather than a prohibition to work around.

Reference and instruction librarians have always mediated between patrons and information sources that carry specific risks. They have taught database searching to users who default to Google, explained the difference between popular and scholarly sources to students who cannot see it, and built citation verification habits in communities where cutting corners is normal. AI in research presents the same professional challenge. It is a powerful tool, widely adopted by patrons before its failure modes are understood, and it creates predictable errors that the library is positioned to prevent.

What is new is not the challenge but the scale and the speed. Students and faculty are adopting AI research workflows faster than institutions can develop formal guidance. Librarians are therefore meeting AI-related research errors in reference consultations, embedded instruction sessions, and research help desk tickets before any institutional response is in place. To address this effectively, librarians need a proactive instruction approach rather than a reactive one. That means teaching appropriate research AI use as part of standard library instruction, not as a specialized add-on for students already in trouble.

The instruction content that matters most at the research support level falls into three categories. First, workflow instruction: explicitly teaching patrons where AI fits and does not fit in the research process, using the workflow map described in the prior section. This works best in the context of a specific assignment rather than as a general AI literacy session, because contextual specificity produces more durable behavior change than abstract instruction. For example, a single-shot session for a research methods course can include five minutes on the appropriate AI workflow for that course's literature review. It addresses this specific assignment, this database, and this citation requirement rather than AI in general.

Second, live demonstration: showing students what AI citation fabrication looks like by trying to locate an AI-generated citation in a database during the session, as described in Module 05. A demonstration shifts behavior more effectively than any verbal explanation, because it makes the abstract failure mode experiential. For the best effect, choose a topic relevant to the course and use a general-purpose AI tool the students already use. That forestalls the dismissive response that a different tool would not make the same errors.

Third, verification as a professional norm: positioning citation verification not as a correction for AI errors but as a professional expectation in all research workflows. For example, framing verification as "the step every professional researcher takes before citing" rather than "the step you need to take because AI is unreliable" produces more sustainable behavior. It connects the practice to professional identity rather than to AI-specific anxiety. The framing is also accurate: professional researchers do verify citations routinely. It positions the librarian as teaching a universal research skill rather than managing an AI problem.

The reference consultation workflow itself can also be updated to integrate AI awareness. When a patron brings a research question, a natural part of the interaction is now to ask: "Have you started working with any AI tools on this project? Let's take a look at what you have." The question normalizes the librarian's role in AI-assisted research without assuming either way about the patron's AI use. It also creates the chance to redirect problematic workflows before they produce academic integrity issues or wasted effort. The library's capacity to provide this consultative guidance in AI-assisted research is one of the strongest arguments for the profession's continued relevance, in an information environment where patrons have increasing access to powerful self-service tools.

Faculty conversations increasingly raise a related but distinct question: whether AI detection tools can identify whether a student's writing was AI generated. ACRL's Knowledge and Understanding competency addresses this directly, noting that AI detection tools are not completely accurate and can be circumvented. To advise faculty responsibly, librarians should be ready to say plainly that a detection tool's report is not evidence in the way a verified citation is. False positives flag legitimate student writing as AI generated. False negatives miss AI-generated text that has been lightly edited. Paraphrasing tools designed specifically to evade detection are widely available. For example, a faculty member who treats a high "AI probability" score as proof of an academic integrity violation is relying on a tool with the same fundamental unreliability this module has described in AI-generated content itself, just pointed in the opposite direction. The library's role is not to validate or dismiss detection tools. It is to bring to them the same calibrated skepticism this curriculum applies to AI output generally. A detection score is useful as one input among several, never sufficient on its own, and never a substitute for the human judgment that an academic integrity determination requires: a conversation with the student, and an examination of their process and drafts.

Research support has acquired a compliance dimension over the past two years that did not exist when most library AI guidance was written. In November 2024, Canada's three federal research funding agencies published tri-agency guidance on the use of generative artificial intelligence in the development and review of research proposals. In July 2025, the United States National Institutes of Health issued a notice addressing originality and the use of AI in research applications. The March 2026 ARL and CNI workshop report treats this as a distinct strategy for libraries. Libraries should develop and maintain authoritative guidance on funder and journal AI disclosure requirements, and embed that guidance where researchers actually encounter it.

This belongs in a library module, not only in a research office, because the question arrives at the library first. A faculty member preparing a proposal asks the subject librarian whether the literature review can be drafted with AI assistance and whether that must be declared. A graduate student asks whether running a draft through a language model counts as assistance requiring disclosure. These are information questions, asked of the person the researcher already trusts with information questions. Answering them with a shrug sends the researcher to a search engine, for an answer that may be a year out of date and written for a different country's funding system.

The most consequential distinction to convey is between using AI to develop your own proposal and using AI while reviewing someone else's. The first is a disclosure question, and the requirements vary by funder, by program, and by the specific role AI played. The second is a confidentiality question, and it is considerably sharper. A proposal or manuscript under peer review is confidential material belonging to someone else. Pasting it into a commercial AI tool transmits it to a third party the author never agreed to, however helpful the resulting summary is to the reviewer. For example, a faculty member who uploads a manuscript under review to summarize it before writing comments has disclosed unpublished research to a vendor. Both the Canadian tri-agency guidance and the NIH notice address this category of use directly. Deleting the conversation afterward does not undo the breach.

To be useful here, the library does not need to become an authority on federal research policy. It needs a current, honest, clearly dated summary of what the major funders relevant to its institution require, with links to the primary notices. It also needs to say plainly where guidance is unsettled rather than filling the gap with a guess. For example, a useful subject guide entry gives the date of the NIH notice and the date of the tri-agency guidance and links both. It separates proposal development from peer review and notes that individual journals set their own disclosure policies. That entry is more useful than a longer document that flattens all of this into a single rule.

The same work extends into research integrity instruction. ARL's companion strategy places library expertise inside Responsible Conduct of Research training and graduate onboarding. There, AI disclosure norms, provenance documentation, and citation of AI contributions can be taught as standard practice rather than handled as an exception at submission time. This material also connects directly to the documentation habits covered in Module 04. A researcher who has recorded which parts of a draft were AI-assisted can answer a disclosure question accurately. One who has not is reconstructing from memory under deadline.

The boundary here is worth stating clearly. The library's role is to know what the requirements are, keep that knowledge current, and make sure researchers encounter it before submission rather than after a problem. What a specific researcher discloses on a specific proposal is for that researcher and their research office to decide. A librarian who issues a definitive compliance ruling on someone else's grant has stepped outside both their expertise and their institutional authority.