BRENNEN

Industries / Government, education and community / Government / Libraries

Cataloguing

Describing and classifying each item, a book, journal or recording, so that people and computers can find it.

7

Milestones achieved, from 1876 to 2023.

3

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1876

    A number for every subject

    Melvil Dewey published a classification scheme, a fixed list of subjects each with its own code, that gives every subject a decimal number so books on the same topic stand together on the shelf. The Library of Congress (LC) has added Dewey numbers to its records since 1930, close to 100,000 a year at its peak in the 1970s and 1980s, and other libraries copy them. Forest Press, the scheme's publisher, became a division of OCLC in 1988.

    oclc.orgloc.gov

    Credit

    Melvil Dewey; maintained today by OCLC

  2. 1901

    Catalogue cards by mail

    On 28 October 1901 LC told about 500 libraries that it would sell them its printed catalogue cards, so a library could buy LC's description of a book instead of writing its own. By the 1960s LC was shipping more than 78 million cards a year, and it printed its last card in 1997.

    LessonCopying one expert description beat having every library write its own, and the idea outlived the card.

    loc.govjournals.ala.org

    Credit

    Library of Congress, under Librarian Herbert Putnam

  3. 1966

    Records computers can read

    MARC (Machine-Readable Cataloging) splits a catalogue record into numbered fields, such as 245 for the title and 100 for the main author, so computers can store and exchange it. In LC's pilot, records went out on magnetic tape to 16 libraries from November 1966, and about 50,000 had been sent by 30 June 1968. MARC became an international standard in 1973.

    eric.ed.govsmecc.orgloc.gov

    Credit

    Henriette Avram, who directed the pilot, and the Library of Congress

  4. 1971

    One library catalogues, all share

    On 26 August 1971 the Ohio College Library Center started an online union catalogue, one shared database of what many libraries hold, so that once one library described a book the others could reuse the record. Cataloguers at Ohio University described 133 books from a single terminal that day; Ohio University calls itself "the first library in the world to catalog online". The catalogue, now WorldCat, held 609,755,407 records and 3,581,565,435 holdings in October 2025, by OCLC's count.

    blog.oclc.orgworldlibraries.dom.eduoclc.org

    Credit

    Frederick G. Kilgour, who founded the organization on 5 July 1967; Ohio University's Alden Library

  5. 2017

    Machines index a nation's books

    Germany's national library, the Deutsche Nationalbibliothek (DNB), began a project in 2010 to assign subject terms by software, and in 2017 made automated indexing its main method, print books included. Staff now add subjects by hand only when the software gives "no or qualitatively insufficient results". Its current system, live since April 2022, is built on Annif, open-source software from the National Library of Finland.

    LessonOne national library's choice spreads to every library that copies its records: a Düsseldorf city library manager wrote that from autumn 2017 children's books and fiction from the DNB no longer came with subject words.

    repository.ifla.orgopus4.kobv.decenl.org

    Credit

    Deutsche Nationalbibliothek; National Library of Finland

  6. 2021

    A heading caught in politics

    Library of Congress Subject Headings (LCSH) is the controlled list of topic terms many libraries use, so that one term finds every book on a topic. On 22 March 2016 LC said it would replace "Illegal aliens", because the phrase "has become pejorative", with "Noncitizens" and "Unauthorized immigration". On 17 May 2016 the House Appropriations Committee voted 25–24 to keep report language telling LC to retain the term, and LC held back. On 12 November 2021 LC replaced the heading with "Noncitizens" and "Illegal immigration".

    LessonSubject headings are public words, and changing one can become a political fight.

    loc.govpbs.orghistorians.orgamericanlibrariesmagazine.org

    Credit

    LC Policy and Standards Division; Dartmouth College, which made the request in 2014

  7. 2023

    AI merges duplicate records

    Duplicate records for the same book clutter search results and split its holdings across records. In August 2023 OCLC began merging WorldCat duplicates with a machine-learning model, trained partly on labels from 336 cataloguers, and it reported more than 9 million duplicates removed by October 2025 across all formats and languages. The merging runs in routine use across the shared catalogue.

    LessonThe counts are OCLC's own and it publishes no error rate; in May 2026 OCLC staff said "we definitely rely on our members to let us know if they see incorrect merges".

    oclc.orgoclc.orghelp-it.oclc.org

    Credit

    OCLC WorldCat Metadata Quality team; volunteer cataloguers

Watch for

  1. LC's machine-drafted MARC records

    LC tested AI models that read e-books and draft MARC records, scored against records its cataloguers had written. Its summary reports about 80% accuracy across all fields, up to 99% on title and author, and below 50% on subject and genre, against LC's own bar of about 95% for full automation. Cataloguers rated the top machine subject suggestion acceptable 42.9% of the time and wrong 29.0% of the time.

    Achieved whenLC uses machine drafts in production and someone outside LC measures them, on a stated sample, at or above LC's 95% threshold for the fields used.

    repository.ifla.orglabs.loc.gov

    Claimed by

    Library of Congress (Caroline Saccucci and Abigail Potter), August 2025

  2. OCLC suggests classes and subjects

    OCLC's cataloguing software now suggests Dewey numbers, LC classification numbers and subject headings drawn from existing WorldCat data, which cataloguers accept or ignore. OCLC says pilot cataloguers reported saving "up to 20 minutes per title", and it gives no accuracy figure.

    Achieved whenan outside library or researcher publishes accuracy and time per title on a defined sample of books.

    oclc.org

    Claimed by

    OCLC, December 2025

  3. Alma's AI record drafts

    Clarivate's Ex Libris AI Metadata Assistant drafts a MARC record in Alma, a widely used library system, from a title and up to four page images. In May 2026 Clarivate added text recognition for multilingual and non-Latin-script books. The University of Liège library, an outside user, reported in July 2025 that it "added valuable content" but that its output must "be checked".

    Achieved whenan outside library publishes measured quality from routine use, such as field accuracy against cataloguers' records or the share of AI-written fields kept after review.

    clarivate.comorbi.uliege.be

    Claimed by

    Clarivate (Ex Libris), May 2026

Open

  1. 01

    Score shipped AI cataloguing tools

    A public test set of 300 print books. For each, four phone photos (title page, the back of the title page, contents page and back cover) are paired with the cataloguer-made MARC record for that edition. The same photos go through Alma's AI Metadata Assistant, OCLC's suggestions, Claude, Gemini, GPT and one local model, and each is scored field by field: ISBN, names, title, edition, publication details, extent, subjects and class numbers, with every heading checked against the authorised form. Publish images only for books whose pages may be shared, such as openly licensed or public-domain books.

    The brief
    Credit · unclaimed
  2. 02

    Audit WorldCat's merged records

    An independent error rate for WorldCat merges. A library takes its own OCLC record numbers, ideally from an export made before August 2023, finds the ones that now point to a different record, and draws a random sample of about 400. A cataloguer compares each pre-merge local record with the surviving WorldCat record and rules the merge correct or wrong, with results broken down by format and language and wrong merges reported to OCLC.

    The brief
    Credit · unclaimed

Open 01 · the brief

Score shipped AI cataloguing tools

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test set of 300 print books. For each, four phone photos (title page, the back of the title page, contents page and back cover) are paired with the cataloguer-made MARC record for that edition. The same photos go through Alma's AI Metadata Assistant, OCLC's suggestions, Claude, Gemini, GPT and one local model, and each is scored field by field: ISBN, names, title, edition, publication details, extent, subjects and class numbers, with every heading checked against the authorised form. Publish images only for books whose pages may be shared, such as openly licensed or public-domain books.

Why it is still open

LC's tests used e-books and LC's internal data. A public set on Hugging Face pairs page images of 9,363 born-digital open-access books with MARC records but reports no model scores, and BiblioPage scores 16 attributes on 2,118 scanned Czech title pages, not MARC fields. OCLC publishes no accuracy figure for its suggestions, and Liège, an Alma library, lists "Evaluate which method is the most effective: AI help or 'classic' manual cataloguing?" as an open question.

huggingface.cogithub.comorbi.uliege.be

Built with what exists

Claude, Gemini and GPT vision models, and a local model through llama.cpp (see the Note "A first stable number for local AI"); Alma and OCLC's tools through a subscribing library; pymarc and MarcEdit to check and score records; LC's id.loc.gov service to check headings.

  1. First step

    Pick 30 openly licensed books the library holds in print, photograph four pages of each, pull their MARC records, and write a pymarc scorer for ISBN, author, title and publication fields.

  2. Done looks like

    a public repository with the images, gold records, prompts and scoring scripts; a results table of exact-match rate and F1 score (a 0-to-1 measure that balances right answers found against wrong answers given) per field per tool; a cataloguer's review of 50 random AI-made records; a second person reruns the scripts and gets the same table.

Open 02 · the brief

Audit WorldCat's merged records

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An independent error rate for WorldCat merges. A library takes its own OCLC record numbers, ideally from an export made before August 2023, finds the ones that now point to a different record, and draws a random sample of about 400. A cataloguer compares each pre-merge local record with the surviving WorldCat record and rules the merge correct or wrong, with results broken down by format and language and wrong merges reported to OCLC.

Why it is still open

OCLC reports more than 9 million AI merges but no error rate, and its staff said in May 2026 that it relies on members to report incorrect merges. The record field that lists merged numbers (MARC 019) does not say whether software or staff did the merge, so an audit has to report all merges in its period.

help-it.oclc.orgoclc.org

Built with what exists

the WorldCat Metadata API, which needs an OCLC cataloguing subscription; pymarc or MarcEdit; a random sampler; Gemini Decision 1 as an optional first pass that rates each pair on fixed questions such as "same edition?" (see the Note "A thousand questions, one pass"). The cataloguer's verdict is the result.

  1. First step

    Export the OCLC numbers from a local catalogue, look them up through the API in batches, and count how many now point to a different record.

  2. Done looks like

    a public report of about 400 judged merges giving the error rate with a 95% confidence interval (about plus or minus 5 percentage points at that size), a list of bad merges reported to OCLC, and a method another library can repeat.