Industries / Government, education and community / Government / Libraries
Digital services
What libraries offer through screens and networks: online catalogues, public computers and internet, scanned books, licensed journals, help by chat, accessible copies for readers who cannot use print, and the security of all of it.
7
Milestones achieved, from 1975 to 2024.
3
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1975
Catalogue moves onto a screen
An online public access catalogue (OPAC) is the card catalogue rebuilt as a database that readers search at a screen. Ohio State University Libraries opened one to the public in 1975 and closed its card catalogues on 1 July 1982, and Dallas Public Library opened one in 1978; the ISKO encyclopedia of knowledge organization lists both among the early projects.
LessonThe first systems copied the card catalogue's author and title lists, and keyword searching came in the 1980s.
Credit
Ohio State University Libraries; Dallas Public Library
1997
Free internet in public libraries
The Bill & Melinda Gates Foundation committed $250 million in 1997 to give US public libraries in poorer areas computers, software and staff training, and by January 2005 it reported more than 47,000 computers installed in nearly 11,000 libraries. In 1996, 28% of library systems offered patrons internet access, by an independent national survey; by 2001 more than 95% of public libraries did, by federal statistics.
LessonBuying machines once was not enough, and the foundation's own 2004 report named keeping access going, starting with hardware and software upgrades, as the challenge.
Credit
Gates Foundation U.S. Library Program; J. C. Bertot, C. R. McClure and D. L. Zweizig (1996 survey); US National Center for Education Statistics
2002
Reference desk goes online
The Library of Congress and OCLC launched QuestionPoint in June 2002, letting librarians answer readers' questions by web chat and email and pass hard ones to other libraries anywhere. More than 300 libraries used it within six months, and about 15% of questions were routed to another library.
LessonA shared service depends on one vendor's plans: in May 2019 OCLC handed QuestionPoint and its round-the-clock cooperative to Springshare, and subscribers moved to Springshare's software.
Credit
Library of Congress, which piloted the idea in 1998 with over 260 libraries; OCLC
2004
Google scans library shelves
On 14 December 2004 Google announced it would scan books at Harvard, Stanford, Michigan, Oxford and the New York Public Library, showing full text of public-domain books and short "snippets" of copyrighted ones. By the time of the ruling it had scanned more than 20 million books. On 16 October 2015 the Second Circuit appeals court ruled that the scanning, the search and the snippets were fair use, and the Supreme Court declined to hear an appeal on 18 April 2016.
LessonCourts accepted copying whole books to make them searchable, as long as readers could not read copyrighted books in full.
Credit
Google and its five partner libraries; Judge Leval wrote the opinion
2016
Accessible copies cross borders
HathiTrust, a shared digital library announced by 13 universities in October 2008, won a June 2014 ruling that it could lawfully offer full-text search and give scanned books to print-disabled readers, people who are blind, have low vision, or cannot hold or read a printed page. The Marrakesh Treaty, adopted at WIPO on 27 June 2013, entered into force on 30 September 2016 with 22 countries, letting approved bodies such as libraries for the blind make braille, audio and large-print copies and send them to other member countries. On that first day Canada's CNIB sent audio books to Vision Australia.
LessonThe need was large: the World Blind Union estimated that fewer than 10% of the million or so books published each year were made available in accessible formats.
Credit
HathiTrust and the University of Michigan Library; Judge Barrington D. Parker; WIPO member states
2023
Ransomware shuts down two libraries
Ransomware is criminal software that encrypts or destroys files so that the attackers can demand payment. On 28 October 2023 it took down systems at both the British Library and Toronto Public Library. The British Library's searchable catalogue was offline until 15 January 2024, its thesis service came back only on 8 July 2026, and its services page still listed research services as affected in October 2026. Toronto refused the ransom and restored public computers on 5 February 2024.
LessonThe British Library's likely way in was a remote-access server without multi-factor authentication, a second proof of identity beyond a password, and both libraries named splitting their networks into separate segments as a fix.
Credit
the British Library's report "Learning lessons from the cyber-attack" (March 2024) and Toronto Public Library's board report (February 2024)
2024
AI answers inside the catalogue
In September 2024 Clarivate released Primo Research Assistant, which answers plain-language questions inside a library's discovery search, the search box that covers its books and licensed articles, by writing each answer from the top five matching abstracts in an index of more than 5 billion records. Four Washington State University librarians published an independent test in December 2025: on 103 searches that had found nothing in ordinary search, 46.3% of the AI tool's sources were on topic, against 45.6% for ordinary search, and its summaries sometimes misstated how a source related to the topic.
LessonDrawing answers from real records did not by itself make results more relevant, so a library should compare an AI tool with its current search before switching it on.
Credit
Clarivate (Ex Libris), for the tool; Blake L. Galbreath, Erica England, Corey M. Johnson and Jen Saulnier Lange, for the test
Watch for
Evidence cards on AI answers
Elsevier says every AI-written insight in LeapSpace, a research workspace built on its Scopus database of over 100 million records, carries a "Trust Card" showing the sources, why each was cited and where research disagrees. The launch release gives no error rates.
Achieved whenan outside group publishes a claim-by-claim test of how often the cited sources support the card and how often a cited item does not exist.
Claimed by
Elsevier, November 2025
Answers tied to the article
ITHAKA, which runs the JSTOR journal archive, says JSTOR's AI research tool uses only the text in front of the reader and shows where in the article each answer comes from. It left beta for all participating institutions in July 2025, and the announcement mentions usability testing but gives no accuracy figures.
Achieved whenan outside test on a public question set reports how often the passages it points to support its answers.
Claimed by
ITHAKA (JSTOR), July 2025
Library scans cleaned for AI
Harvard's Institutional Data Initiative released Institutional Books 1.0: 983,004 public-domain volumes from Harvard's Google scanning, 242 billion tokens (the units AI models count text in), with the original machine-read text and a cleaned version. The initiative reports that the cleaned text scores better on indirect measures, and says those scores are hard to check without the original scans, which it plans to release later. Outside use is real, with 12,455 downloads in one month on Hugging Face, which confirms use, not accuracy.
Achieved whenan outside group measures error rates for both text versions against the page images.
Claimed by
Institutional Data Initiative (Harvard Law School Library), June 2025
Open
01
Reference test for library AI
A public set of 300 reference questions written in the style of real academic and public library chats, without copying any patron's words, each with a librarian-written answer key and a list of acceptable sources. Score library AI tools (Primo Research Assistant, Scopus AI or LeapSpace, Web of Science Research Assistant, JSTOR's tool) and general chatbots (Claude, Gemini, GPT) on whether the answer is right, the share of cited items that do not exist, and the share of citations that support the statement they are attached to.
The briefCredit · unclaimed02
Check OCR against page images
A hand-typed true text for 500 pages drawn from at least 50 Institutional Books volumes across 10 or more languages and three centuries. For each page, measure the character error rate (the share of characters wrong) of the original OCR, the optical character recognition that turned the scans into text, of the initiative's cleaned text, and of a current vision model reading the page image.
The briefCredit · unclaimed
Open 01 · the brief
Reference test for library AI
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of 300 reference questions written in the style of real academic and public library chats, without copying any patron's words, each with a librarian-written answer key and a list of acceptable sources. Score library AI tools (Primo Research Assistant, Scopus AI or LeapSpace, Web of Science Research Assistant, JSTOR's tool) and general chatbots (Claude, Gemini, GPT) on whether the answer is right, the share of cited items that do not exist, and the share of citations that support the statement they are attached to.
Why it is still open
The one independent test of a library AI tool found here, Washington State's of Primo, measured topic relevance at one university, not invented citations, and is not a shared benchmark. A 2025 Rider University study used 28 patron questions and tested only general chatbots. Four searches for a public, librarian-checked question set that scores library AI tools and general chatbots together found none.
Built with what exists
Claude, Gemini and GPT through their programming interfaces; vendor tools through a licensed library account, by hand or with a Browser Use agent where the licence allows (see the Note "Deleting 80% to start over"); the public Crossref and OpenAlex databases to check that cited works exist; XYBench's way of scoring whether an answer finds the reader's real goal, which is what librarians call the reference interview (see the Note "Answering the question behind the question"); Inkling 2 Small as a model said to check its own answers (see the Note "Checked without being asked").
First step
Write 50 questions across five types (known item, topic search, fact, local policy, citation check) with answer keys, run them through two tools to test the grading sheet, and check each vendor's terms before automating anything or publishing scores.
Done looks like
300 questions published under an open licence; two librarians grade every answer separately and agree at a Cohen's kappa of 0.7 or higher; scores per tool for correctness, invented-citation rate and support rate; a script anyone can rerun.
Open 02 · the brief
Check OCR against page images
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A hand-typed true text for 500 pages drawn from at least 50 Institutional Books volumes across 10 or more languages and three centuries. For each page, measure the character error rate (the share of characters wrong) of the original OCR, the optical character recognition that turned the scans into text, of the initiative's cleaned text, and of a current vision model reading the page image.
Why it is still open
The dataset's own report says its text-quality scores cannot be fully checked without the original scans, which it intends to release later, and its Hugging Face listing offers text and metadata, not page images.
Built with what exists
public-domain page images of the same volumes from HathiTrust's full-view collection; Gemini, Claude or GPT vision models to re-read them; an open-source error-rate tool.
First step
Accept the dataset's terms on Hugging Face, draw a random 50 volumes from the metadata, and find their HathiTrust full-view copies.
Done looks like
500 pages, each typed by one person and checked by a second; error rates for each text version with confidence ranges; a script to repeat the test. Publish only page identifiers, your own transcriptions and the scores, because the dataset's terms are noncommercial and bar redistributing its text.