Picking the right tool

When a patron asks a librarian which database to use for a research question, the librarian does not answer by naming the largest or most widely advertised option. The librarian evaluates the patron's topic, the type of sources needed, and the scope of the assignment, and recommends accordingly. For example, a nursing student looking for clinical evidence requires different resources than a history student writing a senior thesis, and a skilled librarian distinguishes between them. The same professional reasoning applies to AI tool selection.

There is no single best AI tool. Some tools perform better for certain task types. Others differ in their privacy implications, the institutional environments they integrate with, and their cost. The landscape also changes rapidly; the strongest option in a category one year may be surpassed the next. To make sound decisions in this shifting landscape, librarians need an evaluation framework, not a fixed list of recommendations. This module provides that framework.

Two dimensions deserve particular weight in a library context: what the tool does with inputs (privacy), and whether it generates or retrieves (reliability). Both affect how the tool can appropriately be used in professional library work. Both are also often poorly understood by librarians who adopt tools based on marketing or peer recommendation alone. The judgment involved in AI tool selection is not different in kind from the judgment librarians have always applied to database selection. It is simply applied to a newer and faster-moving category of resource.

Dakan and Feller name this capacity platform awareness: a working understanding of which AI systems exist, what each does well, and where each is unreliable (Dakan & Feller, AI Fluency: Framework & Foundations, Anthropic Academy, 2025). The concept resists the comparison-chart approach to tool selection. No static list of AI tool ratings stays accurate for more than a few months, and no single evaluation session produces judgment that holds across different task types. What builds platform awareness is repeated use. Over time, you notice that one system handles long documents reliably but loses nuance in multistep reasoning, or that another performs well on structured tasks but drifts when the prompt leaves room for interpretation. These observations accumulate into something more durable than any benchmark score. Dakan and Feller also make a point that holds consistently in library practice: the most effective AI collaborators are domain experts first and AI users second. A librarian who can distinguish a credible source from a plausible-sounding fabrication brings professional judgment to AI output that no amount of tool familiarity can replace. Platform awareness grows on top of existing expertise. It does not replace it, and it is not a prerequisite for beginning.

The following tools are the primary options available to academic librarians as of 2026. Each has genuine strengths and meaningful limitations. The landscape has shifted considerably over the past year: tools that were experimental in 2024 are now production-grade, and new entrants have become impossible to ignore.

ChatGPT (OpenAI) is the most widely recognized AI tool globally and has the largest user community. That means it has the most community resources, tutorials, and library-specific examples available through professional channels. For example, the ACRL and ALA discussion boards have accumulated substantial practitioner experience with ChatGPT that is not yet available for newer tools. The free tier runs OpenAI's current default model (GPT-6 as of October 2026) and is genuinely capable for most drafting and editing tasks. It does limit uploads, deep research, and access to the reasoning models, which becomes noticeable with daily use. OpenAI replaces these models every few months (GPT-4o, the long-time default, was retired from ChatGPT in February 2026). The specific version name therefore matters less than the distinction between the fast "Instant" mode and the slower "Thinking" mode. The paid tier (Plus, approximately $20/month) raises those limits and adds more access to Thinking, which is designed for complex multistep problems rather than drafting. For library work, the default model handles most tasks well. Thinking is valuable for tasks involving complex decision trees or multistep evaluation, such as assessing a vendor contract against a checklist of institutional requirements. On data privacy, inputs on both the free tier and Plus may be used to improve OpenAI's models unless the user opts out in account settings, a step many users have not taken.

Claude (Anthropic) is particularly strong for tasks involving long documents and nuanced, sustained writing. The model family comes in tiers that trade capability for speed: Haiku as the fastest and least expensive, Sonnet as the everyday default, Opus for complex reasoning, and Fable as the most capable, built for long, complex tasks (as of October 2026, Haiku 5.5, Sonnet 5.5, Opus 5.5, and Fable 5.1). The Free plan includes Haiku and Sonnet; Opus and Fable require a paid plan. Claude's context window is among the largest available. A librarian can paste an entire accreditation self-study document and ask Claude to identify gaps relative to a specific standard without hitting length limits that would affect other tools. Claude's Projects feature lets users attach persistent documents, such as a collection development policy, an institutional style guide, and a list of approved databases, that remain available across all conversations within the project. This is particularly useful for librarians who want consistent, institution-specific output across recurring tasks. On data use, a librarian should not assume that paying changes the terms. Since Anthropic's 2025 consumer terms update, conversations on the Free, Pro ($20/month), and Max plans may be used for model training unless the user turns off the "Help improve Claude" setting. Data from users who leave it on is retained for up to five years. Only Claude's institutional plans exclude conversations from training by contract.

Gemini (Google) is most valuable for librarians and institutions whose primary workflow runs through Google Workspace. Gemini integrates directly into Google Docs, Gmail, Drive, and Slides, so it works with documents where they already live rather than requiring copy-and-paste workflows. The current Gemini 3.1 Pro model supports a context window of one million tokens and handles text, images, audio, and video in a single conversation. That is particularly relevant for digital collections librarians working with mixed media. Gemini also uses Google Search as a grounding mechanism, which reduces (but does not eliminate) hallucination on factual questions. For institutions that have signed Google Workspace enterprise agreements, the privacy terms may differ from consumer Gemini; verify this with institutional IT rather than assuming it.

Microsoft Copilot is the most important tool for librarians at institutions running Microsoft 365, and it is substantially underrepresented in library AI discourse relative to its actual relevance. Copilot integrates into Word, Outlook, Teams, Excel, and OneNote. It can draft emails, summarize long documents, and generate meeting notes directly within these applications. For example, a librarian who receives a lengthy vendor proposal as a Word document can ask Copilot to summarize the key terms and flag anything requiring legal review without leaving the application. Institutions with Microsoft 365 education or enterprise agreements typically already include Microsoft 365 Copilot Chat. This web-based assistant is covered by the institution's data protection terms and works with files a user uploads to it. Copilot working directly inside Word, Excel, PowerPoint, and OneNote, however, requires a paid Microsoft 365 Copilot license. The example above therefore depends on which of the two the institution has, which is worth verifying with IT before purchasing a separate AI subscription.

Perplexity is designed to combine AI generation with real-time web search, and every response includes citations to the pages it drew on. For example, a librarian wanting a quick overview of a new federal policy affecting library funding can ask Perplexity and receive a summary with source links to verify. These citations reduce but do not eliminate hallucination risk, since the model may still misrepresent sources it cites. Perplexity is most useful for quick factual orientation, where seeing sources alongside the answer matters more than depth of analysis.

Grok (xAI) is worth knowing about even though it is not yet widely discussed in library professional literature. Grok is built into the X platform and is also available as a standalone product. Its distinguishing feature is real-time access to X platform content, which makes it useful for tracking current professional conversations and emerging library community responses to policy changes. For library practice, Grok is most relevant as a complement for monitoring professional discourse, not as a primary drafting or research tool. Its privacy terms and institutional data governance implications require the same evaluation as any other commercial AI tool before professional use.

Beyond the general-purpose tools, a substantial category of AI products is now built directly into the systems libraries already manage. Over the past year, this category has moved from announcement to general availability across most major platforms. The shift matters for library practice because these tools retrieve from controlled, curated sources rather than generating from a language model's training data. For bibliographic and discovery work, that carries a fundamentally lower hallucination risk than general-purpose AI.

The clearest example is Ex Libris's Primo Research Assistant, now generally available to all Primo customers. It performs retrieval-augmented generation, commonly abbreviated RAG, over the Central Discovery Index, a shared metadata index of roughly five billion records. It returns the five sources most relevant to a patron's query, with citations back to the underlying records. The University of Florida is among the institutions running it in production. This design constrains the tool to recommend sources that actually exist in the index. That is precisely the structural guardrail against fabricated citations that Module 01 identified as missing from general-purpose AI. Clarivate offers a parallel feature for institutions on its Summon discovery platform, the Summon Research Assistant, built on the same retrieval principle.

Cataloging and metadata work has seen comparable movement. Ex Libris has added an AI Metadata Assistant to Alma. On December 8, 2025, OCLC announced AI cataloging features added to WorldShare Record Manager and Connexion, which suggest Dewey and Library of Congress classification numbers and Library of Congress Subject Headings drawn from WorldCat's aggregate data. For example, a cataloger working through a backlog of items with minimal existing metadata can use these suggestions to narrow the search space quickly. Verifying each suggested classification or heading against the item in hand remains the cataloger's responsibility; the suggestion is a starting point, not a substitute for judgment about what the item actually is. In digital collections work, Specto has become a generally available tool for AI-assisted description, organization, and discovery of digitized materials. It extends the same retrieval-grounded approach to special collections and archives.

The database layer has moved as well. ProQuest Research Assistant has been integrated into ProQuest Central since February 2025. Database-level AI assistants are now embedded directly in JSTOR, in EBSCO as AI Insights, in Ebook Central, and in Scopus. Within any of these platforms, the AI assistant operates over that platform's own indexed content. Its summaries and suggested sources are therefore traceable back to records the library already provides access to, rather than to an unknown training corpus.

None of this displaces general-purpose AI from library work; it complements it. ChatGPT, Claude, Gemini, and Copilot remain the tools librarians reach for day to day, for drafting an email, brainstorming a workshop outline, or rewording a policy paragraph. Those are generative tasks, where the lack of grounding in a specific institutional dataset is not a liability. The vendor tools described above earn their place for discovery and metadata work precisely because they are grounded, and at a scale that no individual librarian could review manually: the full Central Discovery Index, the full WorldCat database. What neither category can do is decide, for a specific patron or a specific item, whether the retrieved source actually answers the research question or the suggested subject heading actually describes the item. That determination is a professional judgment, made by a librarian who knows the patron's need or has the item in hand. It remains irreplaceable however good the retrieval becomes.

The free tier of most major AI tools is genuinely useful and a reasonable starting point for a librarian exploring AI for the first time. For example, the free version of ChatGPT can draft a LibGuide introduction, summarize a document, or generate a set of workshop objectives at a quality that is immediately useful for most practitioners. That is substantial capability at no cost, and librarians who have not yet tried these tools have access to more than is commonly assumed.

However, several meaningful differences separate free from paid tiers. Rate limits are the most immediate. Free-tier users encounter usage caps that interrupt workflows during heavy use, particularly near semester peaks or while preparing major instruction programs. Paid tiers remove or substantially raise these limits. Model access is the second difference: free tiers may restrict access to the most capable model version, or provide it on a limited, rotating basis. Paid institutional plans, specifically ChatGPT Business and Enterprise, Claude Team and Enterprise, and Gemini in Google Workspace, also typically include contractual data privacy commitments that free tiers do not. That is a significant consideration for professional library use involving any patron context.

To advocate effectively for institutional AI access, librarians should be ready to make the case in terms administrators understand: time saved on drafting, fewer revision cycles, more responsive patron communication, and clearer data governance compliance. Arguments grounded in workflow efficiency and institutional risk management tend to persuade more than capability demonstrations alone. The data privacy argument is often the one that moves library administration and IT from caution to action: a paid institutional plan provides contractual protection that a free consumer account does not.

Data privacy is the dimension of tool selection that most directly implicates professional library ethics. It deserves more attention than it typically receives in informal librarian conversations about AI. To understand the implications, it helps to know what "using inputs for training" actually means. When a user submits a prompt and receives a response, the tool provider may log both, use them to evaluate model performance, and incorporate them into future training rounds. For example, if a librarian's prompt describes a patron's research question in identifying detail, that information has potentially left the librarian's control. Depending on the tool's terms, it may also have left the institution's jurisdiction.

The general rules as of 2026 are as follows. Consumer accounts, free and paid alike, typically reserve the right to use interaction data for training, with an opt-out available through account settings. These include ChatGPT free and Plus, Claude Free, Pro, and Max, and Gemini on a personal account. The opt-out is a setting the individual user must find and change, not a contractual protection, and paying for a consumer plan does not by itself change the training terms. Paid institutional plans, including ChatGPT Business/Enterprise, Claude Team/Enterprise, and Google Workspace Business/Enterprise with Gemini, typically include contractual commitments not to use customer data for training. Those commitments carry legal weight in ways that opt-out checkboxes in account settings do not.

Data storage and jurisdiction also matter for institutions with specific compliance requirements. For example, some institutions subject to HIPAA, FERPA, or state-level student privacy laws may face restrictions on which AI tools are permissible for certain data types, regardless of the tool's general privacy policy. The appropriate resource for institution-specific guidance is the library's IT department, general counsel, or compliance office, not the AI vendor's marketing materials. Librarians who build working relationships with IT colleagues around AI governance will be better positioned to make responsible tool recommendations than those who navigate privacy questions alone.

AI is inappropriate for certain library tasks, and identifying those tasks clearly is as important as knowing when AI is useful. That clarity is part of what separates professional AI use from uncritical adoption. It is also the practical expression of the skepticism that ACRL identifies as a core competency mindset.

AI is not appropriate in the following situations:

When the task requires verified citations. AI fabricates references, as discussed in Module 01. A librarian should never use AI to find sources; AI may be used to process, summarize, or discuss sources that the librarian has already verified through authoritative channels.

When the input involves patron-identifiable information in an unapproved tool. Patron privacy obligations apply to what a librarian puts into a prompt. A detailed reference question that could identify an individual patron should not go into a free-tier consumer tool. Nor should it go into any tool that the institution has not reviewed and approved for that use case.

When the task requires high-stakes factual accuracy. Statistics, legal information, medical information, accreditation standards, and institutional policy must be verified from primary, authoritative sources rather than accepted from AI output. The same applies to any claim where being wrong has significant professional or institutional consequences. For example, a librarian advising a patron on FMLA eligibility or a faculty member on fair use should consult authoritative legal resources, not an AI summary.

When the patron expects human judgment and empathy. Some reference interactions involve personal circumstances that require discretion, sensitivity, and professional judgment that AI cannot replicate. Examples include a student navigating an academic integrity process, a patron researching a sensitive health situation, or a faculty member in a difficult publication dispute. These interactions require a human librarian. AI may inform the librarian's preparation, but it should not mediate the interaction itself.

When the institution has not approved the tool. Many institutions have developed or are developing AI acceptable use policies. Using an unapproved tool for professional work, even for a seemingly low-risk task, creates compliance and liability exposure that the librarian bears.

New AI tools appear too quickly to evaluate each one thoroughly at launch. To stay current without being overwhelmed, work through a structured set of questions before adopting or recommending any new tool. This makes evaluation repeatable, defensible, and faster than approaching each new tool from scratch.

The questions to apply are as follows. First: who built this tool and what is their business model? Free tools often monetize user data. Tools without a clear business model may disappear or change terms without notice, a meaningful risk for any library workflow that depends on them. Second: what does the privacy policy say, specifically, about training data use, data retention, and data jurisdiction? Vendor privacy pages vary enormously in clarity and specificity, and the details that matter most for library use are rarely in the top-level summary.

Third: is there library community discussion about this tool? The most reliable sources for library-contextualized assessment are LTI (LibTech Insights) from Choice360, Library Technology Reports from ALA TechSource, and the discussion forums of ALA's Core division, including its Artificial Intelligence and Machine Learning in Libraries Interest Group. These sources tend to identify practical limitations that vendor marketing does not mention. Fourth: does my institution have a policy about this tool, or does the acceptable use policy exclude its category? Checking with IT before adopting is far more efficient than explaining an unauthorized tool after the fact.

Fifth: is there a genuine library-specific use case, or is this a general tool being marketed into library contexts without library-relevant features? That kind of marketing is common. Sixth: can I pilot it with low-stakes tasks, such as internal drafts, brainstorming, and non-patron-facing content, before using it for anything that affects professional output or patron services? Piloting with a colleague rather than alone also produces a faster and more reliable assessment, since different task types surface different limitations.

The practical question most librarians face is not which tool is best in the abstract but which tool suits the specific task at hand. For example, the tool best suited for summarizing a forty-page policy document is not necessarily the best one for drafting patron-facing instruction emails. Neither may be the best for a quick factual lookup where source citations matter. Understanding this relationship between task and tool is the applied expression of the professional judgment this module develops.

The following principles guide tool selection by task type. For drafting and editing, including emails, instruction content, LibGuides, policy documents, and committee reports, Claude is generally preferable for longer documents or anything requiring a consistent tone. ChatGPT is useful for shorter tasks, where its large community of examples supports rapid iteration. For tasks requiring Google Workspace integration, such as drafting in Docs, managing in Sheets, and summarizing in Drive, Gemini is the practical choice given its native integration. For work in Microsoft 365, such as generating reports in Word, drafting emails in Outlook, and taking meeting notes in Teams, Copilot is worth exploring even if it is less discussed in library AI circles. Copilot Chat is likely already included in the institutional Microsoft agreement, even where the in-app version requires a separate license.

For factual questions requiring source citations (quick research overviews, policy summaries with verifiable sources), Perplexity is a useful complement to, but not a replacement for, library database search. It is appropriate for orientation to a topic, not for authoritative information. For bibliographic and catalog-level tasks, including discovery search, citation recommendation, cataloging, and digital collections description, the platform-specific tools described above are more appropriate than general-purpose AI. These are Primo Research Assistant, Summon Research Assistant, Alma's AI Metadata Assistant, OCLC's WorldShare and Connexion cataloging suggestions, Specto, and the database-level assistants in ProQuest Central, JSTOR, EBSCO, Ebook Central, and Scopus. They retrieve from controlled sources rather than generating from training data.

To develop genuine judgment about tool selection, deliberately vary the tools you use for low-stakes tasks over several weeks. Note where each tool's outputs differ meaningfully, and build that experiential knowledge before applying it to higher-stakes professional work. This deliberate practice is the most reliable path to tool discernment, the kind of calibrated, task-specific judgment that ACRL subcompetencies 3.4 and 4.5 describe as the professional standard.