Generative-AI systems can produce polished citations for books, articles, journals, government reports and archival records that were never created. People then bring those references to librarians, expecting them to locate the material. The citation may contain a plausible author, publisher, page range or catalogue number, but professional formatting is not evidence that the source exists.
At the Library of Virginia, chief of researcher engagement Sarah Falls estimated that about 15% of emailed reference questions she received were generated by AI. She said some included hallucinated published works and primary-source documents—an estimate for that library’s email questions, not a national statistic. Futurism reported Falls’s account on December 10, 2025.
What librarians are being asked to find
The requests can involve almost any kind of documentary evidence:
- Books with plausible academic titles, publishers and publication dates.
- Journal articles with realistic volume, issue and page details.
- Real scholars credited with articles they never wrote.
- Entire journals or issues that were never published.
- Government reports and institutional publications that were never issued.
- Archival collections, file descriptions and catalogue numbers that sound authentic.
- URLs that lead nowhere, redirect to unrelated pages or reproduce the invented citation.
- Real publications whose author, date, title or pagination has been mixed with invented details.
Those cases are not interchangeable. A real source may be difficult to locate, a real source may be cited incorrectly, or a source may be restricted, unprocessed or held under a different title. Only after those possibilities are considered should a librarian conclude that a citation is fabricated.
Recommended Free Tools
#1 Best Overall
“Not found” is not the same as “nonexistent”
Library catalogues are distributed among institutions, and many collections are not digitized. An item may use a transliterated name, a translated title, an abbreviation or an internal file number. An archive’s public finding aid may be incomplete, while the material itself is restricted or still unprocessed. A work may also have been announced but never published. A failed Google search therefore establishes only that the item was not found in that search.
Why an AI citation can look completely authentic
Language models generate likely sequences of words. They do not automatically query a definitive library catalogue, publisher database or archive for every statement. Bibliographic language is especially predictable: author, title, journal, volume, issue, pages and year form a familiar pattern that a model can reproduce convincingly.
The model may recombine a real author with a nonexistent title, place a real journal around a fictional article, or attach genuine publication details to the wrong work. It can also answer when the appropriate response would be “I cannot verify that.” ChatGPT, Google Gemini and Microsoft Copilot are examples named in the reporting, but the problem is not limited to those products or to one model version.
The International Committee of the Red Cross has warned, as reported by Futurism, that statistical systems can invent catalogue numbers, document descriptions and references to nonexistent platforms when historical records are incomplete or silent. That is particularly risky in archival and humanitarian research, where gaps in the surviving record are normal.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Why proving a false record takes so much work
Producing a catalogue record can demonstrate that a known book exists. Demonstrating that an obscure record does not exist is a different task. Librarians may need to compare several kinds of evidence:
- Union catalogues and national bibliographies.
- Subject databases and specialist indexes.
- Publisher catalogues and journal archives.
- Serials directories and institutional repositories.
- Archival finding aids and local collection databases.
- Direct confirmation from an archivist or special-collections department.
A citation can contain several individually plausible details while being false as a whole. Staff must check variant spellings, translated titles, publication histories and related editions before treating the absence of a result as conclusive. Falls described this problem as especially difficult when a requester claims a unique archival record exists.
Rank #3
How fabricated citations spread
- A chatbot generates a confident-looking reference.
- A student, journalist or researcher copies it into a paper, article, report or reading list.
- Another writer encounters that mention and assumes the source was checked.
- Search engines, scraped pages or AI summaries surface the repeated claim.
- The citation gains apparent credibility through repetition, even though no original record has been produced.
Coverage has described this as a form of citation laundering: a false reference can acquire a trail of secondary mentions and become harder to distinguish from a difficult-to-find genuine source. The Outpost also reported a Chicago Sun-Times freelance reading list in which 10 of 15 recommended books reportedly did not exist. That example is presented here as secondary reporting, not as a universal measure of publishing accuracy.
The archival problem: plausible gaps invite invention
Archives often contain partial evidence. A collection may have missing boxes, incomplete descriptions or periods for which no records survive. Those gaps give a language model room to produce a plausible-sounding bridge: a collection name, a file number and a description that fit the institution’s vocabulary but have no corresponding record.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →The ICRC guidance described in the available reporting directs researchers to consult the organization’s own catalogue and archival resources rather than treating an AI-generated list as authoritative. For any archive, the institution’s finding aid and staff confirmation carry more evidentiary weight than a generated citation.
A verification workflow that works
Ask the system for identifying details, but treat them as clues rather than proof:
- Publisher and publication place.
- ISBN, DOI, ISSN or another stable identifier.
- Journal volume, issue and page range.
- Archive, collection, box and folder information.
- A direct link to the publisher, repository, catalogue or archive.
- The exact page or passage supporting the claim.
Then verify independently:
- Search the exact title in quotation marks. Record spelling and punctuation variants.
- Search author and title separately. This can reveal a real author attached to a different work.
- Check major library catalogues and subject databases. Look for edition, language and transliteration differences.
- Use the publisher’s catalogue or journal archive. Confirm that the issue, pages and author list match.
- Verify a DOI through the DOI registry or the journal site. A DOI that resolves to another article is evidence of a malformed citation.
- For books, compare ISBN metadata, national-library records and the publisher. An ISBN-shaped number alone proves nothing.
- For archival material, search the institution’s official finding aids. If the item remains unclear, contact an archivist with the generated citation attached.
- Compare every field. Check author affiliation, year, volume, issue, pages, publisher and title—not just whether a search result has a similar name.
- Reject weak discovery trails. A result found only on a scraped page, citation farm or AI-generated website remains unverified.
- Disclose the origin. Tell the librarian that the citation came from an AI system so the verification can start with the right assumptions.
What to do when searches fail
Classify the result as unverified while you investigate. Possible explanations include a fabricated citation, a translated or abbreviated title, a misspelled author, a non-digitized item, a local catalogue that search engines cannot index, a subscription database you cannot access, a human-copied error, or a work that was withdrawn or never completed.
Do not convert “I cannot find it” into “it never existed” without checking the relevant institution. Conversely, do not preserve an AI citation merely because a rare archive might theoretically contain it. The burden for using the source in an argument is positive verification.
Free tools Windows power users keep installed
One-click scans. No signup required.
What students and writers should submit
- Do not cite an item you cannot verify in a catalogue, publisher record, archive or database.
- Keep the original AI output so you can explain how the lead was generated.
- Tell an instructor, editor or librarian that the reference was AI-generated.
- Replace it with a source whose record and relevant passage you can inspect.
- Quote or paraphrase only text checked against the original source.
- Do not ask another chatbot to “fix” the citation and then accept the revision without repeating the verification process.
What this means for libraries
Reference staff must balance a welcoming service with finite time. They can help patrons learn search methods, but a request to disprove a unique record may require work across multiple catalogues and an archivist’s specialist knowledge. The practical response is not to dismiss every AI-assisted question. It is to teach the difference between a search lead and documentary evidence, record recurring failure patterns, and correct errors without embarrassing the requester.
Generative AI did not create bad citation practices; copying mistakes, paper mills, predatory publishing and reference-management errors existed before chatbots. Its distinctive effect is to lower the cost and increase the speed and plausibility of producing a complete-looking reference. Verification work then shifts to librarians, teachers, editors, journalists and researchers.
Companies have claimed that newer research-oriented systems hallucinate less, while acknowledging difficulty separating authoritative information from rumors and expressing uncertainty. A lower reported error rate is not a guarantee that any particular citation is genuine. The dependable standard remains an identifiable record and a checkable original.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problems
