Library of Congress shelf-slip, circa 1941, for Arizona: A State Guide: main entry, United States. Work Projects Administration. Federal Writers’ Project. No personal name anywhere on the card.¹ The book in hand runs to five hundred-odd pages of prose — road logs, town histories, essays on irrigation law — and not one of the hundreds of people who wrote that prose appears in the access point a reader would use to find it. The card is not wrong. It is following the rules. That is the interesting part.
The rule in question is main entry — the single access point, usually a personal or corporate name, that a catalog designates as the primary statement of responsibility for a work. Anglo-American cataloging codes spent most of the twentieth century deciding, case by case, who or what could occupy that slot. A corporation could, if the work expressed its collective thought. A person could, if the person had actually written the thing. And a great deal of writing — the WPA guides are only the most famous example — fell into the gap between those two provisions: written by many hands, issued under a corporate imprint, and cataloged under a name that describes the administrative container rather than any of the writing.
The mechanism: an authority file is a theory of personhood
An authority file is the controlled list of authorized forms of names that a catalog uses so that “J. Smith” and “John Smith” resolve to the same person. It sounds like housekeeping. It is actually an ontology. When a cataloger creates an authority record, the file asserts that a name corresponds to a distinct, identifiable entity with a body of work attributable to it. The Library of Congress Name Authority File — built up record by record through the Name Authority Cooperative Program since the late 1970s — is, functionally, a census of people the catalog has agreed to believe in.
The WPA writers were mostly not in that census. The Federal Writers’ Project employed something over six thousand people between 1935 and 1943, among them writers who would later be famous under their own names — Richard Wright, Zora Neale Hurston, John Cheever, May Swenson, Kenneth Rexroth.² For the guides they produced, the cataloging decision was structural: the corporate body took main entry, and the individual contributions dissolved into it. A researcher looking for Hurston’s WPA-era fieldwork has to know to look past the catalog card entirely, into state archives and manuscript collections, because the bibliographic apparatus of the time recorded the employer, not the employed.
This was not malice. It was a rule — a defensible one — meeting a production model the rule’s drafters hadn’t anticipated. And that meeting point, where a rule about naming collides with a text whose authorship is diffuse, delegated, or effectively mechanical, is where catalogs have always had to improvise.
Ghostwriting and the boilerplate shelf
Ghostwriting is the older case, and the catalog’s handling of it has always been quietly pragmatic. Under Anglo-American rules, a ghostwritten celebrity memoir is entered under the celebrity’s name, because the person named on the title page is the person the work is about and the person a reader will look for. The ghostwriter — who did the actual writing — typically appears, if at all, in a note field. The catalog is not recording who wrote the book. It is recording who the book is for retrieval purposes considered to be. Those are different claims, and catalogers have always known it.
Boilerplate pushes the same problem further. Subscription publishers of the nineteenth and early twentieth centuries produced county histories, biographical compendia, and “mug books” assembled largely from pre-written templates with local names inserted — sometimes with the insertion done so mechanically that the same paragraph of pioneer virtue appears in three states.³ The catalog card names the compiler or the publisher. The template’s actual author, whoever first drafted that paragraph, is unrecoverable by design: the production model depended on authorship being invisible. When a genealogist cites one of these volumes, the citation names a person who, in the strict sense, did not write the sentence being cited.
What links the WPA guide, the ghosted memoir, and the mug book is that in each case the catalog made a decision about retrievability and the reader tends to read it as a decision about authorship. The card looks like a statement of fact. It is a statement of policy.
What the record loses
The losses are concrete. Labor historians reconstructing the Federal Writers’ Project have had to rebuild personnel records from payroll ledgers, correspondence, and the occasional memoir, because the bibliographic record — the layer of documentation designed to make works findable — pointed at the agency and away from the workers. The catalog card for Arizona: A State Guide is accurate and, for certain questions, useless. Both things are true at once, and the second follows from the first.
There is also a subtler loss. Authority control creates a feedback loop: names that get authority records accumulate works, which makes them more prominent, which makes them more likely to get future works attributed. Names that never enter the file never start accumulating. The WPA writers who were already established got their project work folded into their records eventually; the ones who weren’t, mostly didn’t. The catalog didn’t erase them, but it also didn’t do the one thing that would have made them findable, and the difference between those two outcomes is invisible unless you go looking.
The acquisitions desk, now
Which brings us to the present, in which the cataloging question is no longer hypothetical. The Authors Guild’s published best practices state flatly that AI-generated text is not copyrightable because it is not “original human authorship,” that such text must be disclosed and disclaimed in copyright registration applications, and that book contracts’ warranty of originality may be breached by undisclosed machine-generated material.⁴ The Guild also recommends that writers hiring ghostwriters or editors contractually require disclosure of any AI use — a clause structure that treats the ghostwriter’s toolkit as something the named author needs to know about, which is the same information the cataloger will eventually need.
An acquisitions librarian now has to decide, for a manuscript that arrives with an unknown proportion of machine-drafted text, what the access point means. The current rules were built for two failure modes: the corporate body and the invisible human hand. A third is arriving — text produced by a system that is neither a person nor an institution with a publications program, whose “authorship” is statistical rather than intentional. The MARC field for main entry does not have a clean answer, and neither does the authority file, because the authority file is a theory of personhood and the system is not a person.
Writers themselves are already living inside this ambiguity. It is now routine for working writers to plan and draft with an AI book generator — a book writing app that produces structured outlines and draft material from a writer’s inputs — and tool makers explicitly frame the output as raw material the writer builds on rather than a finished work.⁵ That framing is a claim about authorship, made by the tool, in the absence of a cataloging rule. The acquisitions librarian is the one who will eventually have to adjudicate it, one access point at a time.
The precedent from the WPA card is not encouraging but it is instructive: when the production model changes faster than the rules, the catalog defaults to the container — the agency, the imprint, the platform — and the hands disappear into it. It took decades of archival labor to recover the names of the Federal Writers’ Project employees from non-bibliographic sources. The equivalent recovery for machine-assisted drafts will depend entirely on what metadata is captured at accession, because there may be no payroll ledger to fall back on.
So the question to carry back to your own institution: when a work arrives whose authorship is distributed across a person, a tool, and a training corpus, what does your access point assert — and would you be comfortable if a researcher, fifty years from now, read it as a statement of fact?
Notes
1. The WPA state guides were consistently cataloged under the corporate heading for the Federal Writers’ Project, e.g., Arizona: A State Guide (New York: Hastings House, 1940), main entry United States. Work Projects Administration. Federal Writers’ Project.
2. Employment figures and participant rosters for the Federal Writers’ Project are documented in Jerre Mangione, The Dream and the Deal: The Federal Writers’ Project, 1935–1943 (1972), and in the National Archives’ administrative records of the WPA.
3. The “mug book” production model — subscription county histories with boilerplate biographical sketches — is described in the bibliographic literature on American county and local history publishing, 1870–1920.
4. The Authors Guild, “AI Best Practices for Authors,” updated May 2026, authorsguild.org.
5. Reedsy’s plot generator documentation, for example, states that the tool “won’t write your novel” and that outputs are “raw material” belonging to the writer, with user inputs pledged not to be used for model training; reedsy.com.