Critical evaluation of AI output

Hallucinated content has a characteristic signature that librarians begin to see reliably once they learn it: it is specific and confident. A general claim such as "AI has transformed academic research workflows" carries low hallucination risk. Specific claims with named details attached carry substantially higher risk. Examples include a citation with journal name, volume, issue, page range, and DOI; a statistic with a precise percentage attributed to a named report; and a quote attributed to a specific person with a publication date. Language models are trained to produce exactly this specificity, because specificity is what confident, authoritative text looks like. The model predicts the tokens that follow a plausible claim, and those tokens often include plausible bibliographic details that do not correspond to real documents.

The practical implication is that the most dangerous AI outputs in a library context are the ones that look the most authoritative. For example, a hallucinated citation does not announce itself as fabricated. It includes a realistic journal name from the appropriate field, a publication year within the expected range, an author name with appropriate disciplinary credentials, and a volume and issue number that fall within the journal's publication history. The DOI may be syntactically correct but resolve to nothing, or resolve to a different article entirely. Such a fabrication is indistinguishable from a real citation until it is verified. That is precisely why verification cannot be optional.

By contrast, certain AI outputs carry lower hallucination risk: structural recommendations, tone and style guidance, brainstorming lists, and summaries of documents the librarian has provided. These are lower-risk categories because they are not factual claims about the world, or because the model is working with provided content rather than generating from training data. To build reliable evaluation instincts, internalize this risk gradient and apply it automatically when reviewing AI outputs.

There is no shortcut to verification, but some strategies make it efficient enough to sustain in daily practice. Librarians do not need to treat every use of AI as a research project. They need a small set of reliable strategies that can be applied quickly to the claim types most likely to be wrong.

The most important strategy is citation verification before claim verification. If AI produces a citation, the first question is not whether the claim it supposedly supports is true. It is whether the cited document exists and says what the AI says it says. This two-step sequence is more efficient than trying to verify the underlying claim by other means. If the citation is fabricated, the claim itself may also be unreliable and should be sourced independently regardless. For example, a librarian who receives an AI-generated paragraph asserting that "Smith and Jones (2021) found that 73% of academic libraries report..." should immediately run that citation through a library database before engaging with the 73% figure at all. If the Smith and Jones article does not exist, the figure requires independent sourcing. If it does exist, consult the article itself to confirm that it actually contains that finding.

The second strategy is asking AI to characterize its own uncertainty, which is more useful than it might first appear. Asking "How confident are you in this specific claim? What are you uncertain about? Where might I find a primary source for this?" will often produce a useful acknowledgment of uncertainty. This is not because the model has a hidden confidence metric. These questions are answered through the same prediction process as everything else, and the model has learned that acknowledging uncertainty fits certain conversational contexts. Treat such responses as leads for further investigation, not as definitive assessments of reliability. They are genuinely useful for identifying which elements of a complex response deserve the most verification effort.

The third strategy is triangulation with authoritative sources. Any factual claim used in professional work (advice to a patron, content in a LibGuide, information in a committee report) should be verified through at least one authoritative source the librarian can cite independently. This is a professional standard, not an optional precaution. For example, if AI reports that a particular database covers a specific range of years or disciplines, the library's own database documentation or the vendor's platform page is the authoritative source. That is where the claim should be confirmed. Triangulation takes additional time. It is the reason AI does not actually speed up research workflows in which accuracy is non-negotiable. What AI speeds up is drafting and thinking, which is a different and genuine value.

Fact-checking AI-generated content is a different activity from verifying a citation, and the distinction matters in practice. Verifying a citation establishes whether a document exists and whether it says what AI claims. Fact-checking establishes whether the underlying assertion is accurate, independently of what AI said and of whether AI cited anything in support. Both activities are necessary in different circumstances, and confusing them leads to verification gaps.

Some factual claims come without citation support: figures, dates, characterizations of policy, descriptions of professional standards. For these, fact-checking begins with identifying what type of authoritative source would be definitive for that category of claim. For example, verify a claim about ALA policy against ALA's own published documents. Verify a claim about ACRL subcompetencies against the ACRL AI Competencies document itself. Verify a claim about federal law against the statutory text or a recognized legal resource. Verify a claim about database coverage against the vendor's documentation. Matching claims to source types is exactly the skill librarians teach students under the heading of source authority, and it applies without modification to AI fact-checking.

The most common fact-checking error among librarians beginning to use AI is using AI to fact-check AI. Asking a different AI tool whether a claim from the first tool is accurate is not verification. For example, asking Claude whether a statistic produced by ChatGPT is correct may produce a confident confirmation, a confident contradiction, or an acknowledgment of uncertainty. None of these is verification, because neither model has access to the authoritative source. Verification requires consulting the primary or authoritative source directly. Web search is not automatically sufficient either. Search results may themselves draw on AI-generated content, or on sources that repeat an inaccurate claim widely enough to make it look credible. To verify AI-generated factual claims reliably, consult the authoritative source, not merely a source that agrees with the AI.

Treating every AI output as requiring full verification is not professionally sustainable. A librarian who must independently verify every sentence of every AI-drafted email will abandon the workflow within a week. The reason is not carelessness; the overhead eliminates the value. Using AI effectively and responsibly requires a calibrated approach that matches verification effort to actual risk rather than applying it uniformly.

This calibration rests on two dimensions: the stakes of the output and the claim type. Stakes are the consequences of an error. A hallucinated figure in an internal planning email to a colleague carries lower stakes than the same figure in a LibGuide used by hundreds of students or in a committee report that will inform a budget decision. The same AI error, in different contexts, requires different verification responses. Claim type, as discussed in the prior section, is the structural risk profile of specific kinds of AI-generated content. Citations, statistics, and quotes carry higher risk than structural recommendations and editing.

Applying both dimensions produces a tiered approach to verification. High-stakes outputs with high-risk claim types require verification of all factual claims through authoritative sources before use. This tier covers any AI-generated content that will be published, presented to administration, used to advise a patron on an important decision, or incorporated into instruction materials. Medium-stakes outputs with mixed claim types include internal planning documents, draft materials that will be reviewed before use, and brainstorming artifacts. They require verification of specific high-risk elements, while lower-risk elements can be accepted provisionally. Low-stakes outputs made up of lower-risk claim types include structural advice for a draft, tone and style suggestions, and reorganization recommendations. These can generally be judged on their professional merits without independent source verification, though the librarian's own judgment remains the final filter.

To develop genuine calibration rather than a mechanical checklist, build a habit: before using AI for any task, identify what category of output is being produced and what an undetected error would cost. This explicit risk assessment takes approximately ten seconds and substantially changes the verification that follows. Communicate the calibration explicitly to colleagues who are newer to AI. Beginners commonly either fully trust or fully distrust AI outputs, and neither instinct produces responsible practice.

Dakan and Feller introduce a three-layer model of evaluation they call discernment, one of the most practically useful frameworks in the AI Fluency course for library professionals (Dakan & Feller, AI Fluency: Framework & Foundations, Anthropic Academy, 2025). The first layer is product discernment: is the output factually accurate, appropriate for the audience, and actually responsive to what was asked? Most librarians begin with this layer, and it is necessary but not sufficient on its own. The second layer is process discernment: how did the AI arrive at this output? This layer matters most in longer or more complex exchanges, where subtle problems accumulate. For example, the AI may quietly introduce an assumption partway through a response and build later reasoning on it. The result can be a plausible-looking conclusion that is wrong from the second paragraph forward. These errors do not surface in a quick read; you have to follow the logic rather than scan the result. The third layer is performance discernment: is the interaction itself working? At a recognizable point in an exchange, refining the prompt stops improving the output. The conversation has gone off track, and starting over is more efficient than continuing to correct. That judgment applies to the interaction as a system, not just to any individual response. Together, these three layers turn critical evaluation from a one-time check at the end into a practice applied throughout, which is where critical thinking belongs.

The ACRL Framework for Information Literacy provides six frames that structure how librarians conceptualize and teach critical engagement with information sources. Each frame connects directly to the evaluation of AI-generated content. For librarians who already teach these frames, that connection offers a ready-made vocabulary for integrating AI evaluation into existing instruction without building a new pedagogical apparatus. The vocabulary is also current. The ACRL Board approved a revised Framework on September 2, 2026, which keeps the same six frames and now addresses AI directly within them. It uses the umbrella term "emerging technologies" for "automated systems which generate or mediate information, including all types of artificial intelligence."

Authority is Constructed and Contextual is the frame most directly applicable to AI evaluation. 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 example, when a student asks AI about the history of a research method, the response synthesizes whatever sources were in the training data, weighted by statistical patterns rather than scholarly standing. The productive question for instruction is not "Is AI authoritative?" but "What would the actual authority be for this claim, and how do I find it?" The revised frame now states this outright: learners "understand that emerging technologies, such as Generative AI, are not inherently authoritative, nor are they capable of critically evaluating information."

Information Creation as a Process frames generative AI particularly well because it foregrounds the production mechanism. Generative AI text comes from statistical prediction, not from research, expert judgment, or evidence evaluation. Students who understand the generation process, even at the conceptual level covered in Module 01, evaluate AI output more accurately than students who treat it as a search result.

Information Has Value connects to AI through questions of attribution and training data. A language model does not cite the sources it learned from, so the value chain of the information it synthesizes (who produced it, under what conditions, with what rights) is entirely obscured. Copyright, attribution, and scholarly credit questions are all embedded in this frame's application to AI. The 2026 revision adds two practices librarians can teach directly. One is giving credit "through proper attribution and citation, including the disclosure of AI use." The other is weighing how AI-generated information differs from information found through human processes in "speed of access, reliability, content quality and discovery processes, and environmental cost."

Searching as Strategic Exploration is the frame most often violated by patrons who use AI as a search tool. A general-purpose chatbot does not search databases; it generates text. A patron who asks it for sources on a topic is not retrieving indexed records from a controlled collection. The model is producing credible-looking citations from statistical patterns. The key distinction for patron instruction is between AI as a thinking tool (appropriate) and AI as a search tool (inappropriate for anything requiring authoritative retrieval). The revised frame does not forbid AI in searching. It treats using AI as a choice the learner makes with "a critical stance," and it lists prompting "by framing, context-setting, iterating, and critically evaluating outputs" as a search skill in its own right. This module's caution is consistent with that. The choice belongs to the patron, and the librarian's job is to make the risks of that choice visible.

Scholarship as Conversation highlights a limitation of AI that students often do not consider: AI cannot represent current scholarly debate. Its training has a cutoff date. Even before that date, its picture of any scholarly conversation reflects what was most statistically prevalent in the training data, not what is most significant in the field. Use AI for synthesis; go to databases to understand current debate. The revised frame also asks learners to "acknowledge the contributing work of others, the involvement and use of AI tools" in their own work. It further asks them to "value human-created sources over artificially generated ones" where original thought and critical research are concerned.

Research as Inquiry most clearly explains why AI does not replace library research. Inquiry involves forming questions, evaluating evidence, reaching conclusions, and revising them, an iterative process of intellectual engagement with sources. AI can support inquiry at specific points, but it cannot conduct it. The revised frame puts it the same way: "because AI outputs are derivative, they cannot replace the critical thinking and human-led inquiry central to research." Instruction that helps students understand this distinction provides durable value however AI tools evolve.

Hallucination refers to AI generating false specific claims. Bias is a different and in some ways more consequential problem. AI systematically overrepresents certain perspectives, populations, languages, and epistemological frameworks while underrepresenting others. These patterns reflect the composition of the training data and the choices made during fine-tuning. To evaluate AI output professionally, a librarian needs to ask two questions. Is a specific claim accurate? And does the overall response reflect a complete and representative view of the relevant knowledge landscape?

The most significant bias pattern in large language models is the overrepresentation of English-language, Western, and American sources and perspectives. Models trained predominantly on English-language web content and books treat English-language norms, examples, and frameworks as the default. For example, a librarian asking AI to describe research library practices will receive a response shaped primarily by practices at American and British research universities. Other models of library practice exist, but they are underrepresented in the training data relative to the volume of English-language library literature. Such a response may be accurate for its context while being significantly misleading about the global landscape of the profession.

A second systematic gap is recency. Models have a training cutoff date, and the cutoff affects reliability differently by subject area. For a domain such as AI itself, where developments occur monthly, a model trained through late 2024 is substantially outdated on its own field. For a domain such as the history of medieval manuscripts, the same cutoff matters much less. To apply this awareness in practice, identify the training cutoff of any AI tool you use and note how quickly the subject area changes. Treat AI outputs on fast-moving topics with proportionally more skepticism.

A third pattern is the overrepresentation of majority views and consensus positions at the expense of legitimate minority scholarly perspectives. Models trained on large corpora tend to converge on the most statistically common framing of a topic. That framing may exclude emerging research, contested findings, or perspectives from communities that publish less. For example, community-based participatory research on a topic may have produced findings that conflict with dominant quantitative studies. A patron researching that topic may receive AI output that represents only the dominant position, because it has more representation in the training data. To counteract these patterns, librarians should treat AI's account of any contested intellectual landscape as a starting point that needs bibliographic expansion, not a complete account.

Recognizing bias in a single AI response, the skill the previous section builds, is necessary but not sufficient for a library that has embedded AI into its systems. When AI shapes cataloging, discovery ranking, recommendations, and patron interactions at scale, bias is no longer a property of individual outputs a librarian happens to read. It becomes a pattern operating across thousands of interactions that no one is reading. ALA's AI guidance responds by asking libraries to move from spotting bias to auditing for it. It calls on library workers to "evaluate AI systems for potential bias in datasets, algorithms, and outcomes before adoption and throughout their lifecycle" and to conduct "regular audits of AI tools used in cataloging, reference services, content recommendation, and patron interactions." The distinction is between an evaluative habit applied by an individual and an institutional practice applied to a system. The second protects the patrons the individual never sees.

Auditing has two moments, and both matter. The first is before adoption: evaluating a tool for bias as a condition of acquiring it, rather than discovering the bias after it is already shaping patron experience. For example, a library considering an AI recommendation feature for its discovery layer can ask the vendor what the system was trained on. It can test the feature against queries in non-English languages and about marginalized communities, and check whether the results keep surfacing the same narrow range of materials, all before signing the contract. The second moment is recurring. A tool that was acceptable at adoption can drift, and bias often becomes visible only at scale, so the guidance asks for regular audits rather than a one-time check. A recurring audit asks three questions. Is the discovery system's ranking quietly reducing the visibility of certain materials? Do AI-suggested subject headings reproduce outdated or offensive terminology for materials about particular communities? Do the recommendations a patron receives narrow rather than widen the range of what the collection offers?

The specific harm these audits are designed to catch is representational. The ALA guidance names it directly, acknowledging that AI systems "underrepresent non-English languages, dialects, and marginalized communities." This harm is invisible in any single interaction and unmistakable in the aggregate. For example, an AI cataloging assistant that performs well on mainstream English-language monographs may systematically mishandle materials in other languages. It may also misdescribe materials documenting Indigenous or immigrant communities. Only an audit that deliberately tests those categories will surface the pattern. To make such audits meaningful, the guidance also calls on libraries to train staff "to recognize and respond to algorithmic bias." An audit is only as good as the person interpreting its results. Recognizing that a ranking pattern disadvantages a particular community is a professional judgment that requires both technical awareness and knowledge of who the collection serves.

Two further commitments from the guidance turn a bias audit from a diagnosis into a practice. The first is that AI-enabled services must always preserve "a clear path to human assistance," so that a patron disadvantaged by an algorithmic pattern is never trapped inside it. The human path is both a service guarantee and a safety valve for bias the audit has not yet caught. The second is that libraries should "use or contribute to open-source AI tools and datasets that explicitly aim to reduce bias and increase representation where they meet operational needs." That points beyond the individual institution toward the collective work of building tools that are more equitable at the source. For example, a library that documents a bias pattern in a vendor tool, and reports it to both the vendor and a professional community, adds to a body of shared evidence that no single audit produces. None of this delegates the underlying professional responsibility to the audit. Whether a tool's biases are tolerable for a specific community, and what to do when they are not, remains the judgment of the librarians who know that community and are accountable to it.

For teaching AI evaluation, demonstration consistently produces more durable learning than instruction. A librarian who spends twenty minutes explaining that AI fabricates citations changes student behavior less than one who spends ten minutes demonstrating the fabrication live and letting students encounter it directly. A demonstration makes the abstract concrete in a way explanation cannot. The resulting skepticism is calibrated and experiential rather than theoretical.

A reliable demonstration sequence proceeds as follows. First, ask the AI tool to identify three peer-reviewed sources on a topic relevant to the course. Choose a topic the students know something about, so they can judge plausibility. Second, attempt to locate each cited article in an appropriate library database while students watch. Third, report the results transparently. Typically, at least one citation does not exist, at least one has bibliographic details that match no real document, and at least one may exist but not say what the AI claims. Fourth, ask students directly: given what you just saw, how does this change how you will use AI in your research process? The question invites students to draw their own conclusions rather than receive a rule, which produces more reliable behavior change. The demonstration takes approximately ten minutes and fits into an existing single-shot instruction session without displacing other content.

At the reference desk, normalization works better than warning. When a patron mentions using AI in their research process, the productive response is not concern but a practical question: "Did you verify those sources? Let's take a look together." This framing positions verification as the expected professional next step, which it is, rather than as a correction of a mistake. For example, a patron who arrives with an AI-generated list of sources can be guided through verification as a reference interaction. That serves both the immediate research need and the patron's long-term information literacy.

The language used in patron instruction also matters considerably. Framing AI as "unreliable" or "dangerous" tends to produce binary responses. Patrons either dismiss the concern and keep using AI uncritically, or overcorrect and avoid it entirely. Neither response reflects the calibrated judgment the librarian is trying to model. A more accurate framing has three parts. AI has specific strengths and specific failure modes. Verification is the professional practice that lets you benefit from the strengths while managing the failure modes. The librarian is the professional most qualified to teach that practice. This framing positions the library's expertise as essential to AI use rather than opposed to it.

The most durable professional practice is one that has been made explicit, documented, and tested, not one held as a vague disposition toward skepticism. To evaluate AI output consistently across the full range of tasks a librarian encounters, develop a personal checklist: a specific, written set of questions applied deliberately to AI outputs before they are used in professional work. A checklist is not bureaucratic overhead. It is the professional equivalent of a pilot's preflight check, turning an evaluative disposition into a reliable procedure.

An effective AI evaluation checklist for library practice should address four categories of questions. The first is output type: What kind of content did AI produce? Does it contain specific factual claims, citations, statistics, quotes, or legal and policy information? The claim types present determine which verification steps apply and how much effort is warranted.

The second category is stakes: Where will this output be used? Is it an internal draft for my own reference, a patron-facing document, a published resource, or a committee report? Higher stakes require more thorough verification. The calibration framework described in the prior section applies here as the decision rule.

The third category is claim-specific verification: For each high-risk claim identified in the first step, what is the authoritative source, and have I consulted it? For citations: does the document exist, and does it say what AI claims? For statistics: what is the primary source, and does the figure appear there? For policy and legal claims: what is the governing document, and does it support the AI's characterization?

The fourth category is gap assessment: What does this response not address that would be relevant to a complete professional answer? AI outputs tend to present confident syntheses that omit contested perspectives, recent developments, and minority scholarly positions. Asking explicitly what is missing is as important as verifying what is present.

To make the checklist actionable rather than aspirational, write it in language that works for your own professional context. Keep it brief enough to complete in under two minutes. Apply it to all AI outputs intended for professional use, not only to outputs that seem suspicious; that consistency is what builds a durable habit. Sharing the checklist with colleagues also creates a common vocabulary for AI evaluation that strengthens the team's practice. It contributes to the kind of institutional AI literacy that ACRL subcompetencies 3.1 through 3.3 describe as the professional standard.

The checklist also has a place in a larger progression. Leo S. Lo's AI Literacy for All: A Universal Framework (2025, preprint) describes Critical Thinking in four levels: a Questioning Mindset, Analytical Evaluation, Critical Investigation, and Informed Critique (Lo, 2025). A personal checklist works at the second level: applying evaluation criteria to AI outputs, detecting potential biases, and assessing the validity of AI-generated content. The bias auditing practice described earlier moves toward the third and fourth levels, examining the assumptions built into AI systems and advocating for critical standards in how they are deployed. The progression shows that individual verification and institutional auditing are not separate skills but stages of the same one.

ACRL's October 2025 AI Competencies for Academic Library Workers includes a framing that practitioners sometimes overlook amid the document's emphasis on building AI fluency: a right to refuse. The framework states plainly that adoption of AI technologies is neither necessary nor beneficial in all cases. It positions that statement as part of AI literacy itself, not as an exception to it. After this module's calibrated skepticism, the framing should feel familiar rather than surprising. The judgment that determines how much verification an AI output needs is the same judgment that sometimes concludes AI should not be used for the task at all. AI literacy, in other words, includes the literacy to decline. A librarian who considers AI for a specific task, finds it unsuitable, and chooses not to use it is exercising the same competency as a librarian who uses AI well. Declining to adopt is not a failure of AI fluency but an expression of it.

The Critical AI Literacy Framework, published in the International Journal of Librarianship in 2025, offers a concrete tool for the evaluation this right to refuse depends on: RACBAC, an acronym for Relevance, Accuracy, Coverage, Bias, Authority, and Currency. Librarians who have taught the CRAAP test for source evaluation will recognize the family resemblance immediately, and that resemblance is the point. RACBAC takes the evaluative habits already built around print and database sources and applies them, term by term, to AI-generated output. Relevance asks whether the response actually addresses the question asked, rather than a related but different question, which AI does with some frequency. Accuracy asks whether the specific factual claims in the response can be verified, the question this module has emphasized throughout. Coverage asks what the response leaves out: which perspectives, time periods, or subtopics are absent. That question connects directly to the earlier discussion of bias and gaps. Bias asks whose framing, language, and assumptions the response reflects by default. Authority asks what gives this response any claim to credibility at all, given that AI itself has none in the traditional sense. Currency asks how the model's training cutoff and the pace of change in the subject area affect the reliability of what it produced.

For example, a librarian evaluating an AI-generated literature overview for a LibGuide can run through RACBAC in the time it takes to read the response once. Does it answer the actual question (Relevance)? Are the cited studies real and correctly characterized (Accuracy)? Does it represent the full range of scholarly positions on the topic or only the dominant one (Coverage)? Does it default to particular regional or disciplinary framings (Bias)? What would actually establish this as a trustworthy overview (Authority)? Is this a fast-moving area where the response may already be outdated (Currency)? This pass takes a few minutes and produces a specific list of what to verify, instead of a vague sense that the response seems fine. RACBAC is a useful starting structure for a librarian building the personal checklist described above. A librarian who already has a checklist will recognize most of RACBAC's questions in it.

What neither RACBAC nor any other framework can do is make the underlying decision. Is a given AI output, however carefully evaluated, good enough for a specific purpose? Or would the task be better served without AI at all? That determination is informed by the framework but not made by it. It remains the librarian's professional judgment, exercised anew for every task, every time.

The right to refuse described in the previous section is ACRL's framing, and it is not the only position the profession has taken. In its March 2026 workshop report with the Coalition for Networked Information, the Association of Research Libraries proposes a strategy on AI limitation and refusal. It affirms one half of ACRL's framing and then draws a line the ACRL document does not. ARL affirms that limiting or declining AI use is sound professional practice in culturally sensitive collections, in interactions with vulnerable populations, and in contexts where human judgment is non-negotiable. It then states that wholesale refusal to engage with AI tools as part of one's professional role is inconsistent with the responsibilities of an information-based profession in the current environment. In ARL's view, the profession's obligation is not uncritical adoption but informed, principled, and accountable engagement.

Practitioners should sit with that tension rather than resolve it too quickly, because both associations are describing something real. ACRL is protecting the individual worker's judgment against institutional pressure to adopt, a pressure most library workers can feel in their own buildings. ARL is describing a professional obligation to understand a technology that now shapes how patrons find and evaluate information, whether or not any individual librarian uses it. Read carefully, the two positions are less opposed than they first appear. Neither association argues for adoption as a default, and neither argues that a librarian must personally use generative AI in daily work.

The distinction that reconciles them is between informed refusal and uninformed avoidance. Informed refusal is what this entire module has been building toward. Consider a librarian who has worked through what a model does, applied RACBAC to its output, and identified the failure modes that make it unsuitable for a given task. If that librarian concludes AI should not be used for the task, they have exercised professional judgment at full strength. That refusal is evidence of AI literacy, exactly as ACRL says. Uninformed avoidance differs in kind rather than degree: it is declining to learn what the technology does. It leaves the practitioner unable to answer a patron's question about AI, unable to evaluate the vendor feature that arrived in the discovery layer, and unable to take part in the institutional conversation about governance. For example, a librarian who has never used a generative AI tool cannot meaningfully advise a student who has used one for two years. Nor can they recognize the failure modes in what that student brings to the reference desk.

This distinction has a practical consequence for how a library treats a colleague who objects. Some objections librarians raise are principled ones about environmental cost, labor, copyright, and the extraction of value from work that was never offered for that purpose. This curriculum addresses those objections elsewhere precisely because they are substantive. A library that treats a principled objection as a communication problem to be managed will handle it worse than one that treats it as a legitimate value question worth engaging. ARL's own field scan also notes that refusal is now a documented and discussed position in the profession rather than an outlier. That is a reason to argue with it seriously rather than dismiss it.

In my own practice, the useful test is a narrow one. Can I explain, to a colleague who disagrees with me, why this particular tool is unsuitable for this particular task, in terms drawn from what the tool actually does rather than from a general position about AI? If I can, the refusal is professional judgment, and I should hold it. If I cannot, what I have is a preference. The honest response is to learn enough to turn it into a judgment, or to abandon it. That test does not tell any librarian what to conclude, and it is not meant to. It only insists that the conclusion be reached rather than assumed, which is the same standard this profession applies to every other source it evaluates.