Library and information science: AI on the institution that organizes human knowledge

Libraries occupy an unusual position in the AI conversation. They have, more than almost any other institutional sector, a long tradition of caring carefully about patron privacy, intellectual freedom, and the responsible handling of information. They have professional ethics that explicitly address these issues. The American Library Association's Code of Ethics prioritizes user privacy. The IFLA Code of Ethics globally takes similar positions.

These professional values map directly onto an architectural argument that favors local inference. Library AI deployments that use cloud LLMs run into immediate tension with the field's own ethical framework. Library AI deployments that use local inference align with the field's values by construction.

The work

Library AI workloads include:

Cataloging and metadata: generating bibliographic records, applying subject headings, classifying materials, producing descriptive metadata for unique collections. The work is fundamental to making collections discoverable.

Reference services: answering patron questions, helping with research, generating reading recommendations, drafting personalized research guides.

Collection development: drafting collection development reports, generating selection recommendations, producing weeding analyses, drafting acquisition justifications.

Discovery services: producing search results, generating subject pathfinders, drafting topic-specific research guides, helping patrons navigate the collection.

Special collections processing: producing finding aids for archives, generating descriptive metadata for unique materials, drafting collection summaries, producing access guides.

Patron services: drafting policy documents, generating patron communications, producing the documentation that library boards and funders expect.

Digital preservation: drafting preservation plans, generating digitization workflow documentation, producing the metadata that long-term preservation requires.

Programming and outreach: drafting program descriptions, generating outreach materials, producing the impact documentation funders and boards expect.

The math

A representative midsize library — academic or public — generates a meaningful AI workload across these functions.

Cataloging volume varies enormously by institution size, from hundreds of items per month at small libraries to tens of thousands at major research libraries. Reference services run continuously. Collection development happens on cycles.

Aggregate AI workload at a midsize library is in the millions of tokens per month. At frontier pricing, the bill is in the low to mid four figures per year.

For major academic and research libraries — the Library of Congress, major university research libraries, large public library systems — the numbers scale to the low to mid five figures per year. For the largest national libraries and major archive systems, the figures climb to six figures.

The cost case is meaningful but secondary to the values and patron privacy arguments that dominate the architecture conversation in this field.

Why libraries are structurally a local-SLM case

The standard properties for local-SLM suitability are present, with the values dimension making the case unusually strong.

The work is narrow within the institution. Each library has its own collection, its own patron community, its own service philosophy, its own classification practices. A model fine-tuned on the library's corpus outperforms a general model.

The work is repetitive in structure. Cataloging records follow predictable formats. Reference responses follow predictable patterns. Collection development analyses follow predictable templates. Specialization compounds.

The patron privacy framework is structural in librarianship. The ALA's privacy principles prohibit libraries from disclosing information about what patrons borrow, read, search for, or research. State library confidentiality statutes in many US states give this principle legal force. The library profession has fought, sometimes in court, to protect these values against subpoena and surveillance pressure.

Cloud LLM use in library workflows runs into immediate tension with this framework. A patron asks a reference question; the question goes to a third-party cloud LLM; the question is now in a third party's logs. The patron didn't consent to that. The library's ethical framework didn't authorize it. The state confidentiality statute may even prohibit it.

The intellectual freedom framework adds another dimension. Libraries protect the right to read what one wants, research what one wants, learn what one wants, without surveillance. The architectural choice is itself an intellectual freedom statement — local inference protects the patron from being known, in a way cloud-LLM deployment cannot.

The professional ethics are explicit. Library professional organizations have begun issuing positions on AI use in libraries. The positions consistently emphasize patron privacy, intellectual freedom, and the importance of the library's own control over the technology.

What changes with local inference

A library AI workflow on a local SLM looks like this.

A model is fine-tuned on the library's corpus — historical catalog records, reference question patterns, collection development documentation, patron-facing materials. The fine-tune captures the library's specific approach and collection focus.

The model runs on infrastructure the library controls — typically on servers in the library's existing technology environment. For academic libraries within universities, the deployment can integrate with the university's broader academic computing environment.

Library work flows through the inference pipeline. Cataloging records get generated, reference questions get answered, collection analyses get produced — all within the library's controlled environment.

The cost flips from per-operation to fixed.

The patron privacy framework is preserved.

The intellectual freedom commitments are made architectural rather than aspirational.

The library's values and the library's technology are aligned.

The values alignment argument

A specific argument for libraries: the values alignment.

The library profession has, for over a century, been explicit about its values — patron privacy, intellectual freedom, free access to information, the public good of knowledge organization. These values aren't marketing claims; they're foundational professional commitments that have been defended in court, in legislature, and in public debate.

The architectural choice between cloud and local AI is, for libraries, also a values choice. Choosing cloud LLMs requires either compromising on the privacy framework or developing complicated contractual workarounds that depend on vendor compliance. Choosing local inference makes the privacy framework structurally guaranteed by the architecture.

For library professionals, this isn't a productivity argument or a cost argument. It's a question of whether the technology the library deploys is consistent with the library's professional identity.

Where the cloud LLM is still acceptable

A narrow set of cases.

For research analytics workflows operating on collection metadata only, without crossing into patron interaction data.

For internal training and continuing education content that doesn't touch patron or collection data.

For some publisher-supplied bibliographic data workflows where the data is already publicly available.

For patron-facing AI work — reference services, discovery, recommendations, personalized research guidance — and for any AI work that touches patron interaction data, the local-SLM case is structural, and the values case may be closer to mandatory.

The pattern, in the institution of knowledge

Avery NXR is not a library tool. It scaffolds Next.js applications. The architectural pattern repeats, with the patron privacy and values alignment dimensions making the case unusually strong in this institutional context.

Library AI is a narrow, repetitive, patron-privacy-protected, intellectual-freedom-relevant workload. The cost case is real but secondary. The values case is what makes the architectural shift not just preferable but structurally aligned with the profession's foundational commitments.

The library AI vendors that build on local infrastructure — with appropriate fine-tuning on each library's collection and service philosophy, integration with the integrated library systems (ILSes), and respect for the field's values — will find willing customers across the institution. The cloud-LLM-default products will face structural tension with the profession's framework that the values dimension of the case makes especially difficult to bridge.

The pattern continues. Libraries are one of the workflows where the architectural choice is itself a values statement. The institutions that move to local inference are aligning their technology with their professional identity. The institutions that don't are accepting tension between the two.