Reference & instruction

Creating a new subject guide from scratch is one of the most time-consuming routine tasks in reference and instruction work. The structural thinking, the description writing, the introduction copy, and the organization of resources into logical tabs are each manageable on their own. Together, they add up to a substantial time investment, often for a guide needed quickly when a faculty member makes a last-minute instruction request. AI is well suited to speeding up the structural and editorial parts of guide creation. The librarian's professional knowledge of the actual resources remains essential and irreplaceable.

The most efficient workflow for guide creation proceeds in stages that become quick once the pattern is established. First, give AI the context it needs to generate a useful structure: the subject area, the course level and institution type, the patron population, and the assignment the guide will support. For example: "I am creating a LibGuide for a sociology research methods course at a community college. Students will be writing an original research proposal with a literature review section. Suggest a logical tab structure for the guide and what each tab should include." This prompt typically produces a reasonable starting structure in seconds. It may not be the final structure, but it is scaffolding to react to and revise. Reacting to a draft is consistently faster than generating structure from scratch.

Second, use AI for description writing once the structure is set. Database descriptions, section introductions, and annotated resource entries are repetitive writing tasks that consume significant time and benefit from AI drafting. For example: "Write a 60-word description of JSTOR for undergraduate sociology students who are searching for peer-reviewed articles for the first time. Plain language, no jargon, emphasize what types of content students will find and how to access it from off campus." This produces usable descriptive text that the librarian reviews and adjusts for accuracy, instead of starting from a blank field. The same prompt structure can then be applied across multiple resources. A forty-five-minute description writing session becomes a ten-minute review and edit.

Third, AI is equally useful for improving existing guides, not only creating new ones. A five-year-old guide may have accumulated outdated language, unclear descriptions, and inconsistent tone. The librarian can paste individual sections into AI and ask for revision: "Here is a section of my library guide on evaluating sources. Rewrite it for first-generation college students who may be unfamiliar with library terminology. Plain language, no jargon, welcoming tone, under 100 words." These revision passes are much faster than rewriting from scratch. They can also be applied selectively to the sections most in need of updating, leaving well-functioning sections unchanged.

The professional boundary that matters is the one between editorial and substantive content. AI can write the introduction to a guide, draft descriptions of resources, and suggest organizational structures. It cannot identify which databases actually suit a given subject area. Nor can it judge whether a database's coverage is current and relevant to the patron population, or verify that the resources it names are available through the institution's subscriptions. That professional judgment is the librarian's contribution. It remains the central, irreplaceable component of guide quality even when AI handles the writing.

Library instruction sessions follow recognizable structural patterns (learning objectives, content sequence, active learning activities, closing assessment) that AI can scaffold reliably for any new context. Generating a lesson plan structure is easy. The challenge is generating one that fits the specific course, assignment, audience, and time constraints a librarian faces. AI adapts to context much better when the librarian provides rich context in the prompt rather than a generic request.

The most useful prompting approach specifies four variables: the patron population and course context, the assignment the session supports, the available time, and the learning outcome the librarian wants to prioritize. For example: "I have 50 minutes with a first-year English composition class whose students are starting their first academic research paper. They will be using Academic Search Complete and the library catalog. The instructor has asked me to focus on finding and evaluating peer-reviewed sources. Generate a lesson plan with a hook, two or three content segments, an active learning activity, and a closing. Active learning should require students to do something, not just watch." This produces a structured session plan that the librarian can review, adjust, and teach. Preparation time no longer goes to structural decisions that could be delegated.

AI is also useful for generating the variety of activity options and assessment questions that make instruction planning genuinely creative. When a librarian asks for five different ways to demonstrate Boolean searching rather than one, the list almost always contains an approach they would not have reached on their own. For example, a prompt such as "Give me five different active learning activities for teaching database Boolean searching to first-year students, ranging from low tech to digital, each taking no more than ten minutes" may produce a mix of familiar approaches and genuinely novel ideas. That variety is particularly useful when planning for different course contexts. The same content delivered to an honors seminar and to a developmental writing course needs very different activity design, and generating options quickly lets the librarian choose appropriately.

AI's ability to help prioritize serves the single-shot instruction challenge well: designing a session that achieves meaningful learning outcomes in a single fifty-minute encounter. A librarian unsure whether to cover database searching, source evaluation, or citation management in a short time slot can ask AI to reason through the tradeoffs: "I have 30 minutes with a criminal justice seminar. The students need to know how to find peer-reviewed sources and how to evaluate them. I cannot cover both well in this time. Which should I prioritize given this population and assignment type, and what is the most efficient way to briefly address the other?" This structured prompt helps the librarian make a principled decision rather than defaulting to whatever was covered last time. For the most useful response, include what students already know and what the assignment actually requires.

AI is also valuable for generating formative assessment questions, discussion prompts, and session evaluations, the parts of instruction design that require creativity but not subject expertise. For example: "Write three exit ticket questions I could use at the end of a single shot session on database searching to assess whether students can identify the difference between a popular and a peer-reviewed source. Questions should take under two minutes to answer and be usable as a quick paper slip." These assessment tools strengthen instruction design without adding preparation time, and the questions can be saved and adapted across multiple sessions.

One of the most time-consuming parts of library instruction work is not creating new content but adapting existing content for new audiences and contexts. A lesson plan developed for a nursing seminar needs substantial revision to serve a history capstone course. A LibGuide written for graduate researchers needs reframing for community college first-year students. A patron FAQ suited to a research university may be entirely unsuitable for a public library community. AI is exceptionally well suited to this adaptation work. A reliable adaptation workflow is among the highest-return AI practices available to instruction librarians.

The adaptation workflow follows a clear prompt structure: provide AI with the existing content, specify the original context, describe the new audience, and give format guidance. For example: "Here is a handout I developed for a graduate research seminar on database searching. Rewrite it for first-generation community college students who have never used an academic database and may be anxious about library research. Same information, accessible language, welcoming tone, no library jargon, under two pages." This produces a genuinely reworked document rather than a superficial revision. The audience specification gives AI what it needs to make meaningful changes to vocabulary, examples, assumed knowledge, and tone.

Reading level adjustment is one of the most practically useful adaptations AI can perform. Some patron populations include English language learners, students with reading disabilities, or community members unfamiliar with academic vocabulary. For them, a librarian can use AI to adjust existing library content to a specified reading level consistently and quickly. For example: "Rewrite this database search tutorial for a sixth grade reading level. Keep all the factual content but simplify vocabulary and sentence structure. Add brief explanations of any library terms that cannot be avoided." The revision preserves the professional content of the original while making it genuinely accessible to a wider patron population. That is a significant service expansion that would otherwise take hours of careful rewriting.

AI is also useful for repackaging the same content in different formats for different delivery contexts. A detailed lesson plan can be condensed into a single-page handout; a handout can be expanded into a slide deck outline; a slide deck can be revised into a self-paced LibGuide tutorial. For example, a librarian with a strong single-shot lesson plan can ask AI for three variants: a condensed version for a fifteen-minute embedded session, a self-paced version for asynchronous delivery through the course management system, and a single-page quick reference card for students to keep. Multiplying formats from a single base document is one of the most efficient uses of AI in instruction work. The core content is developed once, and AI handles the structural adaptation. For librarians whose programs serve many course formats and delivery contexts, this one workflow can cut content development time substantially across a semester.

The ACRL Framework for Information Literacy was first adopted in 2016, before generative AI became a mainstream tool. On September 2, 2026, the ACRL Board approved a revised Framework that keeps the same six frames and writes AI into them. Its introduction notes that "the ubiquity of emerging technologies, such as artificial intelligence (AI), is reshaping how we interact with information," and each frame now carries AI-related knowledge practices and dispositions. AI has not replaced the framework. It has created new occasions for each frame to become visible to students, in ways earlier information environments did not, and the revision gives librarians official language for those occasions.

The Authority is Constructed and Contextual frame is the most directly implicated. AI has no authority in the framework's sense. It does not derive credibility from institutional position, expertise, peer review, or community recognition. It synthesizes patterns from sources that may or may not carry authority, without preserving or communicating the authority of those sources. For instruction, this frame provides the vocabulary for a productive conversation. Rather than telling students that AI is untrustworthy, the librarian can ask them to articulate what makes a source authoritative and then apply that framework to AI output. For example, an activity might ask students to evaluate an AI-generated paragraph using the same authority criteria they would apply to a news article or a database result. Students then identify what is different about a source that has no author, no institution, no publication history, and no peer review.

The Information Creation as a Process frame connects directly to how generative AI produces text. Some students understand, even at the conceptual level covered in Module 01, that AI text comes from statistical prediction rather than research, expert judgment, or evidence evaluation. They evaluate AI output more accurately than students who treat it as a search result. To teach this frame in an AI context, librarians can design activities that make the creation process visible. Students might compare how a Wikipedia article, a peer-reviewed journal article, and an AI response are each produced, and identify what quality controls apply to each. The comparison turns the production process, which is invisible in the final text, into a subject of critical attention.

Searching as Strategic Exploration is the frame most frequently violated by student AI use. Students who ask AI to find sources, rather than using it to develop search strategies, are applying search logic to a generation tool. To teach this frame in an AI context, make the difference between generation and retrieval experiential rather than explanatory. Show students what happens when they ask a chatbot for sources (hallucinated citations). Then show what happens when they search the same topic in a database (actual retrievable documents). This demonstration, which takes five to ten minutes in a single-shot session, produces more durable conceptual change than any amount of verbal explanation. Note that the revised frame does not tell students to avoid AI in searching. It asks that the decision be "informed by an awareness of inherent risks, such as impact on search skill development, the loss of autonomy in shaping the direction of research," and the environmental and societal impacts of AI use. It also counts effective prompting as a search skill. The demonstration above is how a librarian makes those risks concrete enough for a student to weigh.

The Research as Inquiry frame provides useful language for the librarian's role in an AI environment. Inquiry involves forming questions, evaluating evidence, reaching conclusions, and revising them. It is an iterative process of intellectual engagement that AI can support at specific points but cannot conduct on the patron's behalf. For instruction, this frame is most useful for teaching that AI assistance is appropriate when it supports a student's inquiry and inappropriate when it substitutes for it. That distinction, AI as inquiry support versus AI as inquiry replacement, is the conceptual foundation of appropriate research AI use. The ACRL Framework provides the professional vocabulary to teach it clearly. For librarians with embedded instruction and ongoing course relationships, this framing also gives students a durable evaluative principle that applies beyond any single session or assignment.

One further shift is worth naming, because it changes where an instruction session or a reference interaction begins. Clarivate's 2026 survey of more than 500 students, reported in its Pulse of the Library 2026, describes it: 31% start a research project in a general AI assistant such as ChatGPT, Claude, or Gemini, only 10% start at the library website or database, another 9% start in a dedicated AI research tool, and just 1% ask a librarian. The student who eventually arrives is therefore not bringing a question that needs a search built for it. They are bringing an answer that needs interpreting, and the first professional act is reading a text rather than constructing a strategy. Leo S. Lo has proposed a method for exactly that interaction, published in the Journal of Academic Librarianship in 2026: the CARE approach, which stands for Classify, Assess, Review, Enhance. Classify means applying what Lo calls an answer typography. It sorts AI answers into four kinds (factual, interpretive, constructive, and strategic), on the principle that each kind fails differently. Checking a factual answer the way you would check a strategic one catches neither. Review is, in Lo's own description, the load-bearing step, because it checks an answer against a record you actually trust rather than your impression of it. The full method sits behind a subscription and is worth reading before teaching it in detail. The reframing, however, is available immediately and costs nothing to adopt: when a patron opens with an answer, ask what kind of answer it is before asking whether it is correct. What the patron should conclude from that reading remains their decision. The librarian contributes the structure of the conversation, not its verdict.

Faculty are consulting librarians about AI syllabus policies at an increasing rate. This is not always because librarians have positioned themselves as the right resource. AI policy questions sit at the intersection of academic integrity, information literacy, disciplinary norms, and assignment design, which is genuinely library territory. Librarians who develop a clear consultation framework for these conversations will find themselves in a role that strengthens the library's relationship with faculty across disciplines. It also directly serves the institution's need for coherent, pedagogically grounded AI guidance.

The consultation approach that works most consistently begins with the assignment's learning objectives rather than the AI question. Faculty who ask "Should I allow AI in this assignment?" are asking a policy question; what they need is a learning design conversation. To redirect productively, the librarian can open with: "Let's start with what this assignment is meant to develop in your students. Once we understand what learning the assignment is supposed to produce, we can identify what forms of AI use support that learning and what forms undermine it." This reframe shifts the conversation from prohibition versus permission (a false binary) to purposeful design, where the librarian's information literacy expertise applies directly.

AI can assist with the consultation itself. A librarian preparing for a faculty meeting on AI policy can ask Claude for a range of policy examples organized by restriction level, from fully restricted to fully permitted with disclosure required, drawn from the growing body of public higher education policy language. For example: "Generate five examples of AI syllabus policy language representing different levels of restriction. Include a single sentence rationale for each level and note what assignment types each level is most appropriate for." A prepared overview gives the faculty member concrete language to react to, which is far more productive than a blank drafting session. The librarian's role is not to advocate a particular policy position. It is to help the faculty member think clearly about what the assignment's learning goals require and what policy language accurately expresses them.

Librarians can also help faculty think through practical dimensions of AI policy that initial drafts often overlook. How will students document their AI use? What counts as required disclosure? What are the consequences of a violation? How does the policy align with the institution's broader acceptable use standards? For example, a faculty member who writes "AI is not permitted" without specifying what counts as AI use (grammar checkers, autocomplete, translation tools) creates an ambiguous policy that generates more confusion than clarity. The librarian's information and policy background enables exactly this specificity. It improves the policy's practical effectiveness in ways that faculty, expert in their disciplines but not in information policy, may not anticipate.

The library's position in faculty AI policy conversations is strongest when the librarian arrives prepared. That means examples from peer institutions, a working knowledge of the ACRL and ARL guidance documents, and familiarity with what the institution's own AI acceptable use policy requires. Preparation signals that the library is a substantive resource for this conversation, not merely a referral destination. It also builds the kind of faculty partnership that supports instruction programs and collection decisions well beyond the immediate AI policy question.

A librarian at a teaching institution who looks for models in this area will find most published work coming from well-resourced research universities with dedicated AI centers and named directors. That is not useful to someone who is the AI conversation at their college, rather than a member of a team assigned to it. Concrete practice from institutions with the same constraints does exist and is worth knowing. The Iowa Academic Library Alliance maintains a guide that aggregates real documents from its member institutions. These include Des Moines Area Community College's guide to AI and chatbots, its generative AI statement written for faculty to adapt into a syllabus, and the "When is it safe to use ChatGPT?" flowchart from the University of British Columbia that the college uses. Alongside them is Kirkwood Community College's AI literacy tutorial module, built in Brightspace for faculty to copy directly into their own courses. These artifacts are the useful kind. Each one is something a librarian can hand to a faculty member at the end of a fifteen-minute conversation, not a program that would require a position the college does not have. For example, a faculty member who wants an AI policy in their syllabus but does not know what it should say is better served by an adaptable statement another community college has already tested than by a framework document. The librarian who has one saved has turned a vague request into a completed task. What the librarian still supplies is the judgment about which model fits this course, this discipline, and this instructor's actual concern. No shared repository can determine that from the outside.

Patron-facing communication (emails, FAQs, website content, signage, newsletter items, social media posts, handouts) is one of the highest-return, lowest-risk applications of AI in library work. The tasks are repetitive and structurally familiar, the outputs are reviewed before use, and the efficiency gains are immediately visible. These conditions make patron communication the natural starting point for librarians trying AI for the first time. It is also where the professional case for AI use is easiest to demonstrate to skeptical colleagues.

Reference email response is the most immediately useful application for most reference and instruction librarians. A librarian who spends twenty minutes drafting a careful, patron-appropriate response to a common research question can cut that time substantially by having AI draft the response and then editing it for accuracy and personalization. For example: "Draft a friendly, helpful email response to the following reference question from a community college student: [paste question]. The student should be directed to Academic Search Complete for this topic. Warm and accessible tone, under 200 words, offer to follow up if the student needs help." The draft typically needs editing and accuracy review but is structurally sound and covers the essential elements. This is especially useful for questions that require explaining database search to a patron unfamiliar with library resources. AI handles the structural explanation, and the librarian's time goes to the elements that need professional judgment.

FAQ content for library websites and LibGuides is well suited to AI drafting because the format is standardized and the patron population is knowable. For example, a librarian developing FAQ content for a new interlibrary loan service page can ask AI to draft answers to common ILL questions in plain language for undergraduate students. The librarian then reviews the content for accuracy against the library's actual ILL procedures. This workflow produces a complete FAQ draft in much less time than writing from scratch. The essential professional step is verifying every procedural claim, since AI may generate plausible ILL procedures that do not match how the institution's service actually works.

Signage, handouts, and print communication benefit from AI's ability to condense complex information into brief, patron-accessible language. For example: "Write a 50-word sign explaining how to access library databases from off campus, for students who do not know what a proxy server is. Plain language, no jargon, focus on what the student needs to do, not how the technology works." This produces immediately usable sign text that the librarian reviews for accuracy and adjusts for institutional branding. To keep patron communication in the library's voice across channels, give AI a brief institutional description and samples of existing communications. The output will then match the established tone and register more closely. Once a librarian has a prompt that reliably produces communication in the library's voice, it becomes a reusable template for every later task of the same type.

The efficiency gains from AI in reference and instruction work are real. They accumulate most reliably, however, for librarians who build deliberate, systematic practices rather than using AI ad hoc. A librarian who restates their institution and patron population at the start of every AI session is losing half the value that custom instructions or persistent project contexts provide. A librarian who drafts each LibGuide introduction from scratch is not building on patterns that already worked. Realizing the full efficiency requires investment in three dimensions: persistent context, accumulated resources, and shared team knowledge.

Persistent context means setting up AI tools so they know, across every conversation, who the librarian is and what professional context shapes their work. As described in Module 02, custom instructions in ChatGPT, persistent project contexts in Claude, and similar features in other tools let the librarian specify institution type, patron population, preferred tone, and professional norms once rather than in every session. For example, an instruction librarian at a community college might set up custom instructions that include: "I work at a community college with a large first-generation student population. I am an instruction and reference librarian. My content should always use plain language, avoid academic jargon, and assume students may be anxious about library research." This persistent context improves every later output without the librarian re-establishing it each time. It is a one-time investment that pays off in every future session.

Accumulated resources means building a repository of AI-drafted content that has been reviewed, approved, and used: the LibGuide introductions that tested well, the email templates patrons found helpful, the lesson plan elements that worked in specific courses. The repository serves two purposes. It provides ready-made content to adapt for new but similar situations, and it preserves the successful prompts that produced it. For example, a LibGuide introduction that AI drafted and the librarian approved becomes both a usable document and a model prompt. The librarian can return to the original prompt, adjust the course and discipline details, and quickly produce a new introduction for a different guide. Module 10 covers building and maintaining a prompt library systematically. The principle here is that every successful AI output is also a prompt pattern worth saving.

Shared team knowledge is the dimension of sustainable AI practice that individual effort cannot replicate. Effective approaches are most valuable when shared with colleagues, who would otherwise have to discover them independently. For example, three instruction librarians who each develop their own AI practices may arrive at different approaches, with different levels of efficiency, without ever comparing results. A brief team discussion builds a shared vocabulary for AI-assisted instruction work that benefits the whole team. Which prompts have worked for lesson plan development? Which context specifications produce the most consistent output? The most sustainable institutional AI practice is collective rather than individual, documented rather than implicit, and revisited regularly as tools and institutional needs evolve.