AI Is Not the Revolution. Curation Is.

Editorial correction, September 11, 2026. The crawl statistic is no longer used as a proxy for model training. The rural-library example is hypothetical, and categorical capability claims have been replaced with the accountability argument.

AI is not the revolution. Curation is.

Libraries already make consequential decisions about what people can find, trust, and use. AI makes those decisions easier to see because an algorithm now performs some of the selecting and ranking in public view.

That gives librarians a useful place to begin. Before asking whether a tool can generate an answer, ask what it includes, what it leaves out, and who remains accountable for the result. Those are familiar collection-development questions applied to a new system.

Power is the ability to decide what enters someone else's attention

Library science has been working on versions of that question for 150 years.

A card catalog is a curation system. So is a subject heading. So is a display case at the front of the branch. So is a reading list, a readers' advisory session, a collection development policy, a challenge response. Every one of those is a decision about what a person sees, and what they don't.

The internet changed the scale of the problem. Search results, recommendation systems, and generated answers all decide what surfaces. Librarians do not control those systems, but they can still examine their sources, document their limits, choose when to use them, and provide material the systems overlook.

Where this shows up in the AI moment

My concern is whose knowledge a discovery system makes visible. The Common Crawl language distribution describes that crawl, not the training mixture or behavior of any particular model. It cannot establish those model-level claims.

Collection development provides a practical comparison. Librarians evaluate whether sources are reliable, complete, useful to their communities, and worth purchasing or preserving. Using an AI system does not remove those decisions; it adds another source and another set of omissions to evaluate.

The institutional responsibility remains with people who know your community, your collection, your budget, your political context, and the history of who has been systematically excluded from information access and why. A tool’s usefulness does not settle who is accountable for that judgment.

As a hypothetical scenario, picture a small-branch librarian building a collection on reproductive health for a rural county where half the patrons are uninsured and the hospital system's patient education materials are produced by a Catholic health network. An algorithm surfacing "most circulated" or "highest-rated" returns whatever the mainstream produces.

The actual job is knowing which titles to order, which to place near the desk versus spine-out on the shelf, which community organizations to cross-promote, and which gaps in the vendor catalog to fill with interlibrary loan. That local context has to enter the decision somehow. A librarian is responsible for checking it, including when using a model’s suggestions.

Libraries can name that responsibility plainly: access includes decisions about what is collected, described, preserved, and made easy to find. Those decisions can be evaluated against library values such as equity, privacy, intellectual freedom, and accuracy instead of an engagement metric alone.

The roles that actually need to exist

Two familiar library functions become especially relevant here. The first is curatorial: selecting, organizing, and contextualizing collections with judgment about what is reliable and equitable, not only what is popular or highly ranked.

The second concerns rights and privacy: reading the contract, identifying what the library owns or licenses, asking what happens to prompts and patron behavior data, and checking what state confidentiality law protects. AI makes that work more consequential; it does not make it new.

The supply chain problem

In much of the commercial content supply chain, the library appears near the end: Publisher → Distributor → Aggregator → Library → Patron. By the time material reaches the library, other parties have already set many of the prices, formats, licenses, and discovery rules.

Libraries can recover some judgment by participating earlier: setting shared requirements, supporting preservation, improving description, and governing more of the infrastructure that determines what surfaces. The ILS, discovery layer, and ebook platform are parts of that work.

That is the question behind L/30 and MetisLib: whether open, inspectable tools can return specific infrastructure decisions to libraries. They are experiments, not evidence that the larger governance model already works.

Who maintains the map matters

Commercial discovery systems are built around commercial incentives. Libraries have a different responsibility: organize, preserve, and provide access to knowledge, including material that is local, difficult, or unprofitable.

Google, Amazon, and a social-media recommendation feed all select and rank information. The useful question is what each system optimizes for and whether that goal matches the library's.

Library workers can ask the same concrete questions of AI that they ask of a database or ebook contract: What does it contain? What is missing? What does it record? Who can change the terms? Who is accountable when it fails? Those questions keep professional judgment in the room.

Sources

• Common Crawl, language distribution across crawls. In CC-MAIN-2026-34, English is 40.45 percent of pages by primary language and the next largest, Russian, is 6.87 percent. Common Crawl is not any single model's training set, but it is the largest open corpus of the indexed web and the shape of the skew is visible in it.
• American Library Association, "ALA history": the association was founded at a "Convention of Librarians" in Philadelphia, October 4 to 6, 1876. That is where the "150 years" above comes from.

Disclosure: L/30 and MetisLib are mine. Both are open source, neither takes real patron data, and neither is sold to anyone.