AI, labor & the library worker

Labor belongs at the center of any library AI conversation because automation does not distribute its consequences evenly. The people who absorb the cost are rarely the people who make the decision. When a vendor demonstrates a tool that drafts catalog records or answers reference questions, the efficiency is presented as a pure gain. In a staffed institution, however, every efficiency also raises a question: what happens to the hours the tool frees, and to the person who used to fill them? The ALA guidance is explicit that its Labor value "upholds human agency over artificial intelligence and automation." The human capacity to decide, to override, and to remain the author of the work is therefore not a courtesy the technology grants. It is a principle the institution must protect. For example, a cataloging assistant that suggests subject headings is compatible with human agency only when the cataloger keeps full authority to accept, edit, or reject each suggestion. It undermines human agency the moment the workflow makes accepting the suggestion the path of least resistance and questioning it the exception. That distinction is invisible in a vendor demonstration and decisive in daily practice.

The specific danger the guidance identifies is that when a workflow is automated, "it shifts most decision-making to the end of the task," which it describes as "extremely detrimental, especially in areas, such as cataloging, where careful decisions must be made." This is a precise and underappreciated point. When a human catalogs a record from the beginning, the careful decisions happen at the start, informed by the cataloger's knowledge of the collection and its users. When AI produces the record first and a human reviews it at the end, the careful decision moves to a moment when the reviewer is checking a finished artifact rather than composing one. The cognitive posture of checking is systematically weaker than the posture of deciding. For example, a cataloger reviewing a hundred AI-generated records at the end of a workflow is far more likely to approve a plausible error than one who would never have made that error while composing. Reviewing invites acquiescence in a way that authorship does not. This shift is not a failure of individual diligence; it is a structural property of where the decision sits in the workflow.

This is why the labor question cannot be reduced to job counts alone, though job counts matter. The deeper issue is agency. Does the library worker remain the author of the professional judgment, or become the reviewer of a machine's output? Has the institution arranged its workflows to preserve the former, or is it drifting toward the latter? In my practice, the tools that have strengthened my work are the ones I direct and whose output I compose with. The arrangements I am most wary of would reposition me from the person who makes the catalog decision to the person who signs off on it. Judging which arrangement a given tool creates, and whether it preserves the human agency the work requires, belongs to the workers who do the work. It is exactly the judgment this value exists to protect.

The most direct labor harm the ALA guidance addresses is the use of AI "to justify restructuring, job reductions, deskilling, and workload shifts." The library profession's position on this is worth stating without hedging: the guidance opposes it. The concern is not hypothetical. In sectors where automation has arrived, the efficiency narrative has repeatedly turned a tool that could have improved working conditions into a rationale for cutting headcount. Libraries under chronic budget pressure are exactly where an administrator might reach for AI as a reason to leave a vacant position unfilled. The guidance is direct that libraries should not "replace reference, readers' advisory, instructional services, or community support roles with AI recommender systems or AI chatbots," and that "tasks and decisions that require empathy, judgment, and subject-specific knowledge should not be replaced with AI automation." For example, a library that responds to a retirement by deploying an AI chatbot in place of a reference librarian has not achieved an efficiency. It has degraded a service that depends on the human capacities the guidance names, in a way that is hard to reverse once the position is gone.

Deskilling is the quieter and in some ways more corrosive version of the same harm. It does not eliminate the position but hollows out the expertise the position was built on. When AI performs the parts of the work that developed a professional's judgment, and the human is left to review output rather than exercise the skill, the skill atrophies. The profession loses not a job but a competence. The guidance is emphatic that libraries must "preserve core library expertise," naming "cataloging, subject knowledge, access services, instructional design, reference, readers' advisory, and community support." It frames this preservation as an active obligation, not a passive hope. For example, a cataloging department that routes all original cataloging through an AI tool and reserves humans for review will, within a few years, find its catalogers have lost the fluency that made their review valuable. That is how deskilling becomes self-fulfilling. The erosion is slow enough to be invisible in any single quarter and complete enough to be irreversible over a career.

The guidance also insists on a point that reframes the entire efficiency conversation. When efficiency gains are real and verified, they should be used "to improve working conditions and community services, rather than justifying staff reductions." This is the crux. An AI tool that genuinely saves a reference librarian thirty minutes a day has created a choice, not a mandate. The profession's position is that the recovered time should return to the work only humans can do: the deeper research consultation, the instruction session, the collection development that budget pressure has crowded out. It should not become a justification for asking fewer people to do more. In my practice, the honest test of any tool is whether it lets me do the parts of my job that matter more. The honest test of any administration is whether it treats the time a tool saves as capacity to reinvest in the profession or as slack to cut. Library workers must be positioned to make, and to name, that judgment about whether AI is strengthening the work or justifying fewer people to do it.

A recurring failure in institutional AI adoption is that the people whose work will change are the last to be consulted. The ALA guidance treats this as a defect to correct, not an ordinary feature of organizational life. It states that library workers "should have meaningful input into decisions about which AI systems affect their work" and recommends "a staff advisory group, existing staff governance process, or other locally appropriate mechanism to review proposed AI tools." The word meaningful is doing real work in that sentence. A common substitute for genuine consultation, the informational meeting held after the decision is made, satisfies the form of worker involvement while denying its substance. For example, an administration that announces a selected AI cataloging tool and then invites catalogers to a training session has not obtained staff input. It has scheduled compliance, and the catalogers who could have identified the tool's failure modes before purchase were never asked. This sequence is the norm precisely because it is easier. The guidance's insistence on meaningful input argues for the harder and better path.

In practice, meaningful input means a governance structure that involves affected workers before adoption, not after. The guidance points toward a staff advisory group with a real mandate to review proposed AI tools, test them against actual workflows, and recommend adoption, modification, or rejection. It works because the people who do the work know things about it that no vendor demonstration or administrative summary will surface. For example, a reference librarian on such a group will know when a proposed AI tool fails on exactly the messy, half-formed questions students actually ask. That limitation is invisible to anyone evaluating the tool on clean queries. The institution gains this knowledge only if the librarian is in the room before the contract is signed. The guidance also calls on libraries to "identify which work should remain human-led, which work may be assisted by AI." That decision itself should be made with the workers whose expertise defines where the line belongs, not imposed by someone who does not do the work.

Where a library's workers are represented by a bargaining unit, the guidance goes further. When AI systems "materially affect library employment," it says libraries should "consult affected library workers, staff governance bodies, and bargaining units where applicable." This recognizes that decisions with employment consequences belong within the structures workers have built to represent their collective interest. It is not an obstacle to good adoption but a safeguard that makes good adoption more likely. A decision that survives consultation with the people it affects is more durable and better reasoned than one that could only be implemented by avoiding them. In my practice, the AI conversations that went well started with the workers and moved outward. The ones I watched go badly started with a purchase and worked backward to the people. The judgment about which tools serve the work belongs, first and most legitimately, to the people who do it. Building structures that make their voice meaningful is the institutional expression of the human agency this value protects.

The AI tools a library evaluates arrive looking finished and autonomous. Every one of them, however, rests on a foundation of human labor that the product is designed to make invisible, and the ALA guidance asks libraries to look at that foundation before purchasing. Large language models are trained and refined by enormous amounts of human work: labeling data, ranking outputs, and moderating the most disturbing content the model must learn to avoid. This work is often performed under conditions that would trouble any institution committed to labor ethics. For example, reporting by Time magazine in January 2023 documented that workers in Kenya were paid on the order of two dollars an hour to label graphic and traumatic content to make a widely used AI system safer. Scholars including Mary Gray and Siddharth Suri have described such work as "ghost work," because it is essential to the product and erased from its presentation. This labor is the literal substrate of the tools libraries are being sold. By including it, the guidance refuses to pretend the tools appear from nowhere.

The guidance turns this concern into a concrete procurement practice. Before purchasing, it asks libraries to "request documentation about data-labeling and content-moderation labor conditions," to "request documentation about worker protections, compensation, and mental health support for data annotators and content moderators," and to "request supply chain information including training data origins." It states that libraries should "avoid purchasing, renewing, or deploying AI tools when vendors cannot provide sufficient documentation of ethical labor practices." This gives the abstract concern a workable form. The same procurement conversation in which a library asks about privacy and pricing can and should ask who labeled the training data, under what conditions, and with what protections. A vendor's inability or unwillingness to answer is itself informative. For example, a vendor who can discuss model performance in detail but cannot describe the labor conditions behind the training data is revealing a supply chain it has chosen not to examine. The guidance treats that absence as a reason for caution, not a neutral gap.

I want to be honest that this is the labor dimension where library leverage is most limited. A single academic library is a small customer to a large AI vendor and cannot, on its own, reform a global supply chain of data labor. The guidance acknowledges this directly, noting that "individual libraries cannot generate the systemic pressure needed to hold vendors accountable alone." It commits ALA to "direct advocacy with AI vendors and developers" and to coalition work. That does not make the individual library's question pointless. It makes the question part of a larger accumulation of pressure, and it aligns library purchasing with the profession's stated values even where a single purchase cannot change the vendor. For example, when many libraries ask the same labor-ethics questions in procurement, the questions become a market signal that a vendor eventually has to answer. That is how distributed institutional pressure becomes systemic change. In my practice, asking the question has value even when I cannot compel an answer. It establishes that these conditions are visible to the profession and refuses the invisibility the product was designed to maintain. Libraries should insist on judging whether a tool's human supply chain meets the profession's ethical standard, even where that is difficult. The alternative is to benefit from labor conditions the profession would never accept in its own building.

The same dynamic operates inside the profession, at a scale a library can actually see. A peer-reviewed account of a chatbot pilot at the University of Delaware, published in Information Technology and Libraries in 2024 under the title "It Takes a Village," documents what the pilot actually cost in staff time. That included writing and continually revising the response content, monitoring conversations that went wrong, and coordinating across units in ways no one had budgeted for. The authors state the general principle plainly: chatbots are designed to hide the effort of the people behind them, that labor is substantial, and it needs to be recognized. For example, a chatbot presented to administration as a way to handle routine questions without staff time will, in its first year, consume more staff time than the questions did. A library that did not name that cost in advance has no way to argue for the position it now needs. ALA's final guidance also gives library workers language for when the argument turns from workload to headcount. It states that libraries should not use AI to justify staff reductions, deskilling, or replacement of professional judgment. It also establishes a right to explanation: if AI was used to make a decision that affected a patron, that patron has the right to know why. This language is a professional standard rather than a bargaining position, which is precisely what makes it usable in a room where the library worker is outranked. An efficiency calculation does not answer what a library owes the people who do its work, and the profession has now written down that it will not be answered that way.

There is a form of labor in the AI economy that libraries are positioned not only to witness but to enable. The ALA guidance names it with unusual directness: the patron contributions used to train and improve AI systems are, in its words, "often uncompensated, uncredited, and extracted as free labor." This reframing is worth sitting with, because it identifies as labor something usually described in the softer language of data. A patron's searches, questions, reading choices, and interactions with a library system may be fed into a vendor's model to make it more capable. If so, the patron has performed work, generating the human behavior the model learns from. That work has been taken without payment, credit, or, in most cases, awareness. For example, a discovery system that improves its recommendations by learning from every patron's search behavior is being trained by the unpaid labor of the library's community. The value that labor creates goes to the vendor, not to the patrons who produced it. This arrangement is easy to overlook precisely because the extraction is frictionless and the contribution invisible.

The library's role in this is not incidental. The library is often the institution that collects the patron behavior and the party that decides whether a vendor may use it. That is what makes the labor framing actionable rather than merely descriptive. The guidance's no-harm principle insists that libraries "ensure patron data is not used to train AI models without consent." This is the library's point of leverage, because the library administers the relationship in which that consent is or is not obtained. For example, a library reviewing a discovery vendor's contract may learn that patron queries feed the vendor's model training. It has then discovered a labor extraction it is positioned to refuse. It can require that the feature be disabled, that the data be excluded from training, or that patrons get a genuine opportunity to opt in rather than a default that silently opts them in. Such a refusal exercises a responsibility the labor framing makes visible. The library is not only protecting privacy in the abstract but declining to conscript its community into unpaid work for a commercial model.

The connection to the profession's older commitments is close: the labor framing and the privacy framing point to the same protective action from different angles. A library that keeps patron reading histories and reference interactions out of AI training systems is protecting patron privacy and refusing to extract patron labor at the same time. The two obligations reinforce each other. In my practice, thinking of patron data as labor rather than only as information has sharpened how I read vendor contracts. It turns an abstract data-flow clause into a concrete question about whose work is being taken and who benefits. The library actually holds the judgment about whether its community's behavior may be extracted as unpaid training labor, through the contracts it signs and the defaults it accepts. Exercising that judgment on the community's behalf is a direct expression of the stewardship the profession owes the people it serves.

The labor value would be an empty gesture if it stopped at prohibition, and the ALA guidance does not. It specifies the affirmative obligations a library takes on when it adopts AI in ways that affect the people who work there. The first is that libraries must "provide paid time and institutional support for workers to learn how AI functions" and "how AI ethics apply to library work." This recognizes a reality that unfunded expectations obscure. Learning to use AI tools well, and to understand their ethical dimensions, is work. Requiring staff to acquire that competence on their own time is itself a labor harm dressed up as professional development. For example, an administration that expects reference staff to become proficient with an AI tool but schedules no paid time to learn it has quietly shifted the cost of adoption onto the workers. They must either absorb it as unpaid effort or fall short of an expectation the institution created. This arrangement is common, and the guidance's insistence on paid learning time directly corrects it.

The second affirmative obligation concerns transitions. AI adoption sometimes genuinely changes what a job consists of, and the guidance requires that libraries "establish retraining, transition support, or other locally appropriate supports before using AI to substantially change or eliminate job tasks." The word before is essential. Retraining offered after a role has been hollowed out is a severance formality. Retraining offered before the change, as preparation for a redefined role the worker will still hold, honors the institution's obligation to the people whose work it is reshaping. For example, a technical services department introducing AI-assisted metadata workflows honors this obligation when it retrains staff for the higher-judgment review and exception-handling roles the new workflow requires. It violates the obligation when it changes the work first and addresses the human consequences afterward. This sequencing is the practical difference between adoption that carries workers forward and adoption that leaves them behind, and the institution controls it entirely.

Beneath both obligations lies the principle the guidance returns to repeatedly. Verified efficiency gains should be used "to improve working conditions and community services, rather than justifying staff reductions," and libraries should "assess labor impacts before adopting, renewing, expanding, or reconfiguring AI systems." Here the module's argument becomes a practice a library can actually adopt, because it prescribes a sequence. Assess the labor impact before adopting, and involve the affected workers in that assessment. Fund the learning the adoption requires, and retrain before roles change. Treat any efficiency the tool produces as capacity to reinvest in the work and the workers, not as a reason to reduce them. For example, a library that runs this sequence around an AI cataloging tool ends up with catalogers who were consulted and trained on paid time. They move into roles that use their judgment more fully and are freed for the collection work that mattered but never had hours. That is the version of AI adoption the profession should want. In my practice, the presence or absence of this sequence is the clearest signal of whether an institution is using AI to strengthen its people or diminish them. Library workers are entitled to make that judgment and insist on it. Human agency over automation, the principle this entire value defends, means nothing if the humans it protects are not the ones deciding how the automation is used.