AI Essentials for University Librarians: Six Skills Your Patrons Already Need

Patrons are already arriving at your library through ChatGPT and Perplexity, whether you have prepared for it or not. A 27-month analytics study at Drexel University found that ChatGPT alone drove roughly 85% of AI-mediated referrals to library systems by October 2025, with Perplexity, Gemini, and Copilot sending the rest. The discovery pathway that once routed every patron through your catalog now often starts inside a chatbot, and the citations those patrons bring to your reference desk arrive pre-fabricated. The question is no longer whether to engage with AI, but how to engage with it well enough that your service to faculty and students gets stronger, not weaker.

This tutorial walks through six skills the Association of College & Research Libraries (ACRL) AI Competencies (October 2025) and the broader framework landscape actually ask of you: what AI is and why it is now a discovery pathway, how to write prompts that return useful answers, how to evaluate AI tools for accuracy and accessibility, how to teach AI literacy in three different pedagogical shapes, how to support research when patrons arrive with AI-fabricated citations, and how to anchor your own ongoing learning in ACRL × UNESCO × ARL. Every chapter closes on what you can offer your patrons next.

Why AI Is Now Part of Every Librarian's Job, and Where the Standards Point

Let us start with the practical reason this matters: AI search engines are sending measurable traffic to academic library collections. Kim and Stanislaw's Drexel study tracked 27 months of Google Analytics 4 data, from August 2023 through October 2025, and counted 20 distinct AI platforms referring users to library systems. The first such referral landed on August 25, 2023, from Perplexity. By October 2025, the study recorded 871 users in a single month arriving via AI. The discovery pathway that used to begin at the library catalog now often begins inside a chatbot, and the librarian's role is shifting accordingly KimStanislaw2026.

What you are actually engaging with is a large language model (LLM), a text-producing system that responds to prompts. The way it works can be visualized as a loop: the prompt you type goes in, the model processes it, and an answer comes out. The Prompt Report, a systematic survey of prompt-engineering techniques covering 1,565 papers, describes this as a three-step cycle: run an inference, evaluate the result, modify the prompt and try again Schulhoff2024. This same loop will reappear in every chapter, because every skill we discuss is a way of operating that loop more deliberately.

Caption: The prompt-engineering process loop, showing the three repeated steps the Prompt Report identifies: run an inference on your prompt, evaluate whether the output meets your need, and modify the prompt template before the next round. An "extractor" step cleans the model's natural-language reply into a usable answer.

Standards you can cite when colleagues ask "who says we need to do this?"

Three professional bodies have published AI guidance for libraries, and each speaks to a different audience. ACRL (October 2025) is the librarian-facing competency standard: it tells you personally what you need to be able to do. The ARL Guiding Principles for Artificial Intelligence (April 2024) is the institutional-strategy voice: it tells research libraries how to position themselves on AI. UNESCO's AI Competency Frameworks (August 2024) is the educational-system vocabulary: it gives you language for negotiating with faculty and administration who speak the broader campus curriculum ALA-ACRL2025.

The useful way to think about this, if you already think in catalog terms, is: ACRL is your catalog, the vocabulary you personally use at the reference desk. ARL is the collection-development policy, what your institution commits to. UNESCO is the shared classification scheme, the language everyone across campus speaks when they discuss AI literacy. The frameworks reinforce each other rather than compete.

Caption: The ACRL AI Competencies, visualized as a 2×2 grid of four categories (Ethical Considerations; Knowledge & Understanding; Analysis & Evaluation; Use & Application) under a horizontal strip of five Guiding Mindsets. This is a worker-generated summary of the official ACRL document, not an infographic published by ACRL itself.

A note on that last image: the four-category diagram is a summary derived from the official ACRL document. ACRL has not released a published infographic, so for the authoritative text and bullet-level sub-competencies, consult ala.org/acrl/standards/ai directly.

What the standards say your patrons already need

ACRL's Ethical Considerations category (sub-competencies 1.1–1.5) asks you to address fairness, privacy, intellectual property, accountability, and transparency in AI use. The Knowledge & Understanding category asks you to grasp how AI tools work well enough to explain them to a first-year student. Analysis & Evaluation asks you to assess tools for reliability and bias. Use & Application asks you to embed AI into your professional practice, including sub-competency 4.3: developing effective prompting strategies for optimal AI output ALA-ACRL2025. The chapter that follows unpacks that competency first.

By the end of this tutorial, you will be able to map any patron request onto one of those four ACRL categories. That is the promise chapter 1 makes; chapter 6 redeems it.

Writing Prompts That Actually Return Useful Answers

The first time you ask a chatbot to draft a reference-desk reply, the answer will likely come back generic. The second time, if you have told it to act as a reference librarian and given it context, it will come back more useful. The third time, if you have shown it two example answers before asking for the new one, it will sound like something you would actually send. That progression, from generic to tailored, is what prompt engineering means in practice. The Prompt Report defines prompt engineering as the iterative discipline of refining a prompt template over time, backed by a taxonomy of 58 techniques organized into six categories: Zero-Shot, Few-Shot, Thought Generation, Ensembling, Self-Criticism, and Decomposition Schulhoff2024.

Caption: A network view of seven interconnected areas within the prompting literature. The image signals that prompt engineering is a structured academic discipline with a comprehensive surveyed taxonomy, not folk wisdom you can replace with a single clever sentence.

Three techniques that cover most library reference work

You will not use all 58 techniques in a week. Three cover the bulk of reference-desk and instruction scenarios:

1. Zero-shot with role and context. You give the model a role ("act as a reference librarian"), a context ("a third-year biology student needs peer-reviewed sources on CRISPR ethics"), and a format ("return three database recommendations in bullet form"). The model has no examples to learn from; it must infer what you want. This works well for straightforward drafting tasks.

2. Few-shot, also called in-context learning. You show the model two or three example question-and-answer pairs before asking the real question. The GPT-3 paper introduced this paradigm, and it remains the textbook technique whenever you want output in a specific style Schulhoff2024. For library work, this is how you get a chatbot to write in your library's voice or to format reference-desk replies consistently.

   

   Caption: The Brown et al. (2020) in-context learning paradigm. The model sees a sequence of paired translation examples, then encounters a new prompt ("cheese →") and completes it in the same pattern. Few-shot prompting works the same way: examples precede the real question.

3. Chain-of-thought. You ask the model to reason step by step before giving its final answer. For library work, this is useful when you ask the model to build a search strategy ("think through which databases would be best for this topic, then suggest three searches"). The intermediate reasoning tends to surface assumptions you can correct.

The missing piece: the iterate-evaluate-modify loop

Most patrons skip this step. They type a prompt, read the first answer, and accept it. The Prompt Report's process loop shows why that fails: you are supposed to evaluate the output, modify the prompt, and run another inference. The first answer is a draft, not a deliverable Schulhoff2024. You may need two to four rounds before the answer is actually fit for purpose.

Think of it as iterative search refinement. You would never type one keyword into a database and accept the first ten results; you would add field limiters, try synonyms, and scan the results for new vocabulary. Prompt refinement follows the same shape.

A reference-desk template you can paste tomorrow

Here is a template that combines role, context, few-shot, and format directive, the four ingredients the Prompt Report highlights as best practice. Paste it into ChatGPT, Copilot, or Claude, fill in the bracketed sections, and adjust.

Role: You are a reference librarian at a university library.

Context: A [year-level] [major] student asks: "[patron's actual question]."

Examples:
Q: Where can I find peer-reviewed sources on climate adaptation policy?
A: Try Web of Science for citation tracking, JSTOR for policy archives, and
   PAIS Index for public-policy grey literature. Limit to 2018–present.

Q: How do I cite a dataset in APA?
A: Use the author of the dataset, the year of publication, the title in
   italics, the publisher/repository, and the URL.

Now answer: [patron's actual question]

This template satisfies ACRL sub-competency 4.3, "Develop effective prompting strategies for optimal AI output" ALA-ACRL2025.

Caption: The ACRL Use & Application category, with sub-competency 4.3 (prompting) highlighted. This is a worker-generated diagram from the ACRL document, not an official ACRL infographic.

The takeaway for this chapter: prompt engineering is a discipline, not a one-shot typing exercise. The iterate-evaluate-modify loop is what makes it engineering.

Evaluating AI Tools: Accuracy, Bias, Privacy, Accessibility

The patron who arrives at your reference desk with an AI-generated bibliography has likely been handed something that looks scholarly but is not. A February 2025 benchmark tested eight free-tier chatbots (ChatGPT, Claude, Copilot, DeepSeek, Gemini, Grok, Le Chat, and Perplexity) on 400 reference requests across five academic disciplines. Only 26.5% of the references returned were fully correct, 33.8% were partially correct, and 39.8% were erroneous or fabricated. About four out of every ten references a student receives from these chatbots are wrong or invented CabezasClavijo2025.

Caption: Eight chatbots ranked by share of completely correct (green), partially correct (purple), and wrong-or-fabricated (red) references. Grok and DeepSeek produced the most correct answers; Copilot produced none, every one of its references was fabricated or wrong.

The chatbot-to-chatbot spread is wide. Copilot fabricated every single reference in the benchmark, while Grok produced no fabricated references at all on the test set. The implication for your reference work is that the choice of chatbot changes the prior probability that a citation is real, but it does not eliminate the verification step for any of them. The error-rate distribution shown below makes the same point a different way: the average of "about two errors per reference" hides the fact that some chatbots cluster their errors at the high end (Copilot, Le Chat) and others at the low end (Grok, DeepSeek) CabezasClavijo2025.

Caption: Errors per reference by chatbot, showing how the average conceals real differences between systems. Copilot and Le Chat cluster at the high end; Grok and DeepSeek cluster at the low end.

The four criteria you should evaluate any AI tool against

Librarians evaluating AI tools need a defensible, framework-anchored checklist. The ARL Guiding Principles and the ACRL Ethical Considerations sub-competencies converge on four dimensions ARL2024.

Accuracy is the criterion the chatbot benchmark speaks to directly. Bias is a documented concern, and ARL Principle 2 asks librarians to recognize that AI systems encode the biases of their training data. Privacy and copyright sit at the institutional-policy level: ARL Principle 5 reminds libraries that patron data flowing through third-party AI tools is subject to that third party's terms, and Principle 6 points to the Library Copyright Alliance Principles (July 2023) for guidance on AI-generated works ARL2024.

Accessibility deserves equal billing

Evaluating AI tools for accessibility means asking whether the output is reachable by patrons who use screen readers, who need captions on video, who navigate by keyboard, or who require alternative formats. A four-week evaluation at one Indian university found that integrating AI-driven library services improved accessibility task-completion from 62% to 88%, and WCAG 2.1 compliance from 71% to 93%, alongside engagement and satisfaction gains KumarP2025. A complementary study in special libraries cataloged an eight-category taxonomy of AI-powered assistive technologies: optical character recognition, text-to-speech, NLP-based voice search, image recognition for visual content, automated accessibility evaluation, personalized recommendations, monitoring and feedback, and compliance auditing PaulChauhan2024.

Both studies have limits: the first is a single-site, four-week window with novelty-effect risk; the second is a 20-patron qualitative case study. Treat the numbers as evidence that AI tools can improve accessibility measurably, not as a guarantee that every AI tool does so by default.

Caption: The ARL Guiding Principles (April 2024), visualized as a four-quadrant poster grouping the seven principles into thematic clusters. Worker-generated from the ARL document text; not an official ARL infographic.

What this lets you offer your patrons

You now have a four-dimension evaluation checklist (accuracy, bias, privacy/copyright, accessibility) backed by the ARL Principles and ACRL Ethical Considerations. When a faculty member asks "is this AI tool safe for my students?", you can answer with criteria rather than vibes. And when a student hands you an AI-generated bibliography, you can tell them with numbers that about four of every ten citations will need verification before they can trust it.

Teaching AI Literacy: Three Pedagogical Shapes You Can Adapt

Students are already using AI for research. Most have not been taught how to evaluate what AI gives them. That gap is your teaching opportunity, and it slots naturally into the information-literacy instruction you already do. ACRL's Framework for Information Literacy, the older six-frame standard (Authority is Constructed, Information Creation, Information Has Value, Research as Inquiry, Scholarship as Conversation, Searching as Strategic Exploration), already accommodates AI as a coding scheme, not as a new subject KoChiu2024.

Three pedagogical shapes triangulate the same ACRL-grounded content. Each is legitimate; each has limits.

Shape 1: coding-scheme LibGuide design

Ko and Chiu's content analysis of 28 ACRL Excellence Award winners' AI LibGuides scored coverage against 16 sub-themes spread across the six ACRL IL frames. The most-covered cells were the mechanisms of AI (frame 2-1), community support (5-2), training-process issues (2-2), different AI models (2-3), and the value of AI in academic work (5-1). The least-covered cells were AI for searching (6-3), search strategies with AI (6-2), prompt-dependent answer changes (1-2), and research as inquiry applied to AI (4-2). If you are auditing or building an AI LibGuide, the four under-covered cells are where you can add the most value KoChiu2024.

Shape 2: active-learning RPG quest

Chapman University's Lorekeeper's Trial embeds AI literacy in a required First-Year Foundations Canvas module using a four-role RPG structure GreeneDechow2026.

Caption: The four Lorekeeper roles and their assigned research tasks. Ranger hunts the stacks; Cleric evaluates a news article; Rogue selects peer-reviewed articles; Wizard uses AI to refine the question. AI is one bounded role, not a search-replacement.

The lesson the four roles teach is structural: AI is one role in a research process, not a search-replacement. The Cleric still has to evaluate a news article; the Rogue still has to select peer-reviewed sources. The Wizard uses AI only to refine the question. No measured learning gains are reported, so treat the design as illustrative of the pedagogical structure rather than evidence of effectiveness.

Shape 3: tool-forward four-workshop series

Smith's "AI Literacy Is Information Literacy" series at a small private four-year university runs four 50-minute workshops Smith2026.

Smith is candid: faculty skepticism and slower-than-hoped workshop uptake are real. That honesty is useful because it tells you the obstacle is institutional, not pedagogical.

Choosing the shape that fits your teaching context

If you already teach one-shot sessions and want a defensible structure for an AI LibGuide, start with the coding scheme. If you have access to a required first-year course and want to prove AI literacy can be taught without making it a product demo, the RPG structure is worth studying. If you have faculty partners willing to schedule workshops, the four-workshop series is copy-able. All three shapes map onto ACRL's four competency categories, so whatever shape you choose can be measured against the standard ALA-ACRL2025.

Supporting Research When Patrons Arrive via ChatGPT

When a student opens ChatGPT, asks for "five peer-reviewed sources on urban heat islands," and walks into your library with a citation list, they have just skipped your discovery layer entirely. The retrieval happened upstream, in a chatbot they did not necessarily tell you about, and the chatbot has handed them a list that may include real journal articles, partial citations, and outright fabrications in roughly equal measure. Your job, as the librarian in the loop, is to make sure the patron does not submit those citations without verification.

The empirical case for why this matters comes from Drexel again: across 27 months, AI platforms drove a growing share of sessions to library online systems, and the Research Repository, especially Electronic Theses and Dissertations (ETDs), absorbed the largest share of those sessions KimStanislaw2026.

Caption: Monthly AI-mediated user traffic to Drexel library systems, August 2023 through October 2025. The surge from under ten users in late 2023 to 871 users in October 2025, with ChatGPT dominating, is the empirical reason AI-enhanced research support is now a chapter rather than a footnote. Traffic surges correlate with the October 2024 launch of ChatGPT search features and the August 2025 release of GPT-5; the authors note this is correlation, not causation.

This is the same work reference librarians have always done when a patron arrives with a citation from a less-reliable upstream source, a predatory journal, a bad database export, an undergraduate rumor. The discovery half of the workflow has changed; the verification half has not. ARL Principle 4, "no human, no AI," captures the requirement that a human remains accountable for AI-assisted decisions ARL2024.

A librarian-in-the-loop research-support protocol

Here is the operational shape of that loop, drawn from the practitioner strategies identified by Nsirim's survey of 97 Nigerian librarians and triangulated with the ARL Principles and the ACRL Ethical Considerations Nsirim2025.

1. Discovery upstream. The patron finds your content via ChatGPT, Perplexity, or Gemini, and lands on a library record (most often an ETD, a journal article, or a repository item) without ever using your discovery layer.
2. Citation check. For every AI-generated citation, the patron or librarian verifies it against CrossRef, OpenAlex, the catalog, or Google Scholar. The chatbot benchmark tells you to do this for every reference: about four in ten will be wrong or fabricated.
3. Source evaluation. The librarian applies the ARL Principles 2 and 3 lens (bias awareness and transparency) to whatever the AI surfaced. Does the source actually exist? Does it say what the AI claims it says?
4. Citation work. The patron cites the verified source correctly, including any required conventions for acknowledging AI assistance in the research process.

Each step is something you can teach the patron to do themselves. That is the offering you can extend: a drop-in "verify your AI citations" clinic, a course-integrated "AI-assisted lit-review" session, or simply a LibGuide page that walks through these four steps.

What the numbers say to a faculty member

When a faculty partner asks whether students should be allowed to use ChatGPT for literature reviews, you can now answer with both standards and evidence. ARL Principle 4 says the human stays accountable. ACRL sub-competency 1.4 says you are responsible for the accuracy of AI-assisted work. The chatbot benchmark gives you the probability that any single citation is wrong: roughly 40% CabezasClavijo2025. And the traffic data shows this is already happening at peer institutions. You are not preparing for a hypothetical future; you are catching up to a present that arrived in late 2023.

Your Roadmap: ACRL × UNESCO × ARL, on One Page

You now have six skills: a mental model of how LLMs respond to prompts, a prompt-engineering discipline with three core techniques, a four-criterion evaluation framework, three pedagogical shapes for teaching AI literacy, a librarian-in-the-loop research-support protocol, and the empirical awareness that patrons are arriving via AI. What remains is to anchor your ongoing professional development in the frameworks that already exist, so your learning is institutionally supported rather than ad hoc.

Three frameworks matter most for North American academic librarians. ACRL (October 2025) is the librarian-facing competency standard: five Guiding Mindsets (Curiosity, Skepticism, Judgment, Responsibility, Collaboration) and four competency categories (Ethical Considerations; Knowledge & Understanding; Analysis & Evaluation; Use & Application) ALA-ACRL2025. UNESCO's AI Competency Frameworks (August 2024) are the educational-system vocabulary: 15 teacher competencies in a 5×3 matrix (Human-centred mindset; Ethics of AI; AI foundations and applications; AI pedagogy; AI for professional learning) crossed with progression levels (Acquire, Deepen, Create), and a parallel 12-competency students framework UNESCO2024. ARL (April 2024) is the institutional-strategy voice: seven Guiding Principles, with Principle 4 ("no human, no AI") as the operational anchor for the research-support chapter in this tutorial ARL2024.

The three frameworks open with the same disposition: a human-centred mindset. ACRL's five Guiding Mindsets; UNESCO's first row in both matrices; ARL's Principle 4. That convergence is not coincidence. It is the consensus that AI literacy has reached as a defined professional field, and your job is to translate that consensus into your institution's vocabulary.

A one-page personal-roadmap template

Pick one ACRL category to focus on this semester. Find the UNESCO row that maps onto it. Pair it with one ARL principle. Schedule one concrete classroom trial.

Fill in the last column with something specific to your teaching context. A LibGuide revision, a workshop, a reference-desk script, a liaison conversation with a faculty partner. The frameworks give you the column headers; your schedule supplies the content.

Transferability beyond academia

If you doubt this model works outside research libraries, consider the Bord Bia Copilot programme. Bord Bia is Ireland's state food board, and a librarian-led three-phase Microsoft Copilot upskilling programme (Self-Directed Foundations → Peer-to-Peer Knowledge Exchange → Creative Finale) trained non-academic staff in AI use McKeownLombard2026.

Seven design principles guided the programme: library-led, plain English, time-bound, learning-by-doing, community, psychological safety, and ethics-first. Six guardrails set the boundaries: protect sensitive information, assume errors will occur, verify outputs, keep human accountability, be transparent about AI use, and use AI for options rather than authority. The RGCSE prompt framework (Role, Goal, Context, Source, Expectations) is a librarian-friendly mnemonic you can pair with the Prompt Report's 58-technique taxonomy for everyday use McKeownLombard2026.

The structural insight is this: ACRL × UNESCO × ARL is your anchor for the next five years of professional development. Not a checklist to complete, but a map to navigate by. Patrons are arriving via AI. The librarian's role is to engage with it well enough that service gets stronger, not weaker.

Next steps you can take this week

You are not behind. The frameworks were finalized in 2024 and 2025; the empirical traffic data documents AI-mediated discovery already arriving at one academic library, with patterns consistent enough that peer institutions should expect similar dynamics. The library is positioned to be the campus authority on AI use precisely because you already teach source evaluation, citation work, and information literacy. AI literacy is information literacy. The skills you built this week are the same skills you have always had, extended to a new information environment.

References

1. ACRL AI Competencies for Academic Library Workers — Approved by the ACRL Board of Directors, October 2025 — https://www.ala.org/acrl/standards/ai
2. ARL Research Libraries Guiding Principles for Artificial Intelligence — Association of Research Libraries, April 2024 — https://www.arl.org/resources/research-libraries-guiding-principles-for-artificial-intelligence/
3. UNESCO AI Competency Frameworks for Teachers and Students — UNESCO, August 2024 — https://www.unesco.org/en/articles/ai-competency-framework-teachers
4. Kim, H. M., & Stanislaw, S. (2026). Understanding Generative AI-mediated User Engagement with Academic Library Resources — arXiv:2607.20328 (accepted for College & Research Libraries, May 2027) — https://arxiv.org/abs/2607.20328
5. Cabezas-Clavijo, Á., & Sidorenko-Bautista, P. (2025). Assessing the performance of 8 AI chatbots in bibliographic reference retrieval — arXiv:2505.18059 — https://arxiv.org/abs/2505.18059
6. Schulhoff, S., et al. (2024). The Prompt Report: A Systematic Survey of Prompt Engineering Techniques — arXiv:2406.06608, v6 February 2025 — https://arxiv.org/abs/2406.06608
7. Greene, E., & Dechow, J. (2026). The Lorekeeper's Trial: Teaching AI Literacy through Active Learning in the Library Classroom — CCLI 2026 conference paper, Chapman University — https://digitalcommons.chapman.edu/librarian_articles/86
8. Ko, M., & Chiu, M. (2024). How Can Academic Librarians Support Generative AI Literacy: An Analysis of Library Guides Using the ACRL Information Literacy Framework — ASIS&T 2024 poster — https://doi.org/10.1002/pra2.1159
9. Nsirim, J. (2025). Hallucinations in Artificial Intelligence and Human Misinformation: Librarians' Perspectives on Implications for Scholarly Publication — Folia Toruniensia — https://doi.org/10.12775/ft.2025.004
10. Kumar P, U., & Srinivasulu, M. (2025). AI-Driven Library Services: Assessing User Engagement, Satisfaction, and Accessibility Impacts — Asian Journal of Applied Communication and Computing Methodology — https://ajaccm.com/journal/index.php/ajaccm/article/view/142
11. Paul, S., & Chauhan, S. (2024). Enhancing Accessibility in Special Libraries: AI-Powered Assistive Technologies for Patrons with Disabilities — arXiv:2411.06970 — https://arxiv.org/abs/2411.06970
12. Mc Keown, P., & Lombard, E. (2026). Building AI Literacy Through Library-Led Copilot Training: A Practice-Based Case Study — Health and Social Care in the Community Journal, Summer 2026 — https://journals.ucc.ie/index.php/hslj/article/view/4871
13. Smith, A. (2026). AI Literacy Is Information Literacy: One Academic Library's Plan for AI Instruction — Information Matters, January 12, 2026 — https://informationmatters.org/2026/01/ai-literacy-is-information-literacy-one-academic-librarys-plan-for-ai-instruction/