The Paper Exists. It Just Doesn't Say That.

In October, a peer-reviewed paper about Z-Library appeared in New Review of Academic Librarianship, a Taylor & Francis journal. One of its references was a 2021 TorrentFreak article called "Publishers continue crackdown on shadow libraries." TorrentFreak has never published an article with that title. When its reporters went through the whole reference list, they had trouble finding at least ten of the 44 references.

The missing references were not gibberish. A 2019 article in the Harvard Journal of Law & Technology was credited to a real IP law professor, but the cited issue's table of contents does not include it, and the cited pages hold two articles on other topics. A 2019 article in College & Research Libraries was credited to a real researcher who did co-author a 2019 article on a similar subject, in a different journal and under a different title. The paper's acknowledgments say Consensus AI was used to sift the literature and ChatGPT to improve the language. TorrentFreak is careful to add that this "doesn't prove that AI wrote these references." Taylor & Francis told the site it had referred the concerns to its ethics teams.

One clause in that report is easy to read past. Besides the references it could not find, TorrentFreak noted "others that didn't necessarily back the stated claims."

That second group is what this article is about. Someone on X, agreeing that checking sources against Semantic Scholar is the right layer, named what that layer leaves open:

"The gap I'd still watch is quote-level: the paper exists, but it doesn't say what the sentence claims. Does the review panel check the cited passage itself, or only that the citation resolves?"

— @KeyoAPI, October 23, 2026

The audits, tools and penalties described below are aimed at the first question: does the cited work exist? The second question is whether the work says what you cite it for. A lookup cannot answer that one.

Two failures that look the same in a reference list ​

A reference that points to nothing ​

This is the failure that gets counted. A team of six researchers (Zhao, Wang, Stuart, De Vaan, Ginsparg and Yin) audited 111 million references across 2.5 million papers in arXiv, bioRxiv, SSRN and PubMed Central. They report "a sharp rise in non-existent references following widespread LLM adoption," with a conservative estimate of 146,932 hallucinated citations in 2025 alone. Two more of their findings matter here. First, the fake references "disproportionately assign credit to already prominent and male scholars." Second, preprint moderation and journal publication "capture only a fraction of these errors." Nature's news coverage led with the finding that the rate was highest on SSRN, a preprint server that mostly holds social-science research.

A separate audit in The Lancet (Topaz and colleagues) looked at biomedical papers. As STAT reported it, the team found around 4,000 fabricated citations in 2,800 papers. In 2023, 1 in 2,828 papers had at least one fabricated reference. By 2025 it was 1 in 458, and in the first seven weeks of 2026 it was 1 in 277. The study began with something that happened to its lead author. He had used an AI chatbot to help edit an editorial and checked the citations. After several rounds of edits, the journal's editor still flagged one that was wrong. "I was deeply embarrassed," he told STAT. "I checked for that, and it still almost happened to me."

Institutions are reacting. In May, Nature reported that arXiv would ban researchers from posting for one year if a submission contains references hallucinated by AI tools, or other "incontrovertible" signs that AI output was not carefully checked. In October, two researchers at Estonian universities (Pisek and Veromann) wrote to Nature arguing that AI-hallucinated citations should be treated as research misconduct. Researchers rank it high too. Ithaka S+R surveyed 4,003 US researchers in late 2025 and asked each to pick the five most severe forms of research misconduct from a list. Seventy-one percent included "citing false sources." Only misrepresenting findings (84 percent) and plagiarism (82 percent) were chosen more often.

A reference that points to the wrong thing ​

The second failure barely shows up in those numbers, because a lookup cannot find it. When the Lancet audit came out, Retraction Watch asked Mohammad Hosseini, who studies research integrity and AI ethics at Northwestern. He called the study "low-hanging fruit" and the "tip of the iceberg." The site summarized his view this way: the "bigger and more important problem" is citations that aren't wholly hallucinated but are inaccurate, biased or incomplete. "We are far from being able to even detect them or do anything about them," he said.

This failure is older than chatbots. A 2017 review in PLOS ONE (Mogull) pooled 15 studies of medical research articles and recalculated the rate of quotation errors at 14.5 percent of the citations examined. Most of those errors were major ones, where the referenced source "either fails to substantiate, is unrelated to, or contradicts the assertion." The review counted improper secondary citations separately, at 10.4 percent.

In STAT, Hosseini described what has changed:

"Previously you would read a paper and then make notes on it. Once you were writing a paper, you would be like, 'Is that paper or book relevant? Is that paper not?' It was a much more reflective process, whereas now people simply use their hunches to prompt ChatGPT or other AI tools, and then they have a bunch of citations that they can sprinkle over their papers, and that is not a healthy practice."

The Z-Library paper contains both kinds. Some of its references point to nothing, even when the author and the journal are real. Others, in TorrentFreak's words, "didn't necessarily back the stated claims." As one researcher on X put it: "Hallucinated citations have ruined so many literature reviews already."

That is what the reporting says. The rest of this article is my reading of it.

A lookup checks only one end of the chain ​

Hosseini's description is a chain with three links. You read a passage. You write something down about it. Later, while writing, you decide whether it belongs in the argument. The citation is the visible end of that chain. A hallucinated reference is what you get when the visible end is produced without the other two.

Reference checkers work on the visible end. Given a title, authors and perhaps a DOI, they ask a registry whether a matching record exists. That is a useful question, and it catches a TorrentFreak article that was never published. It does not reach the passage. A real paper, by the author you named, in the journal you named, passes the lookup whether or not its page 12 says what your sentence says.

The finding about credit makes this harder. My reading is this: if fabricated references tend to go to names that are already well known, a fake reference and a misused one look alike. Both look like the citation a reviewer expects to see. A lookup clears the misused one, because the paper is real. Only a reader opening the page would catch it.

"Not found" is not the same as "fake" ​

Lookups also fail in the other direction. Renee Hoch, head of publication ethics at PLOS, told STAT what the publisher saw while piloting reference checkers:

"we have seen a lot of false positives in which legitimate references are flagged due to erroneous, incomplete or inaccurate information in reference lists, formatting issues, the source material's publication language, or limitations of commonly used research literature databases"

This falls hardest on bibliographies that registries cover least well: a monograph from 1987, a regional archive, a journal that publishes in Lithuanian, a modern edition of a classical text. Many of these have no DOI, and some are in no database at all. For a bibliography like that, "no record found" says little either way, and the real check happens at a library shelf. My guess is that this is part of why fields built on print and archival sources tend to find these errors late. A tool that treats every unmatched reference as suspect trains you to ignore it. One that treats every unmatched reference as fine misses the fabricated ones.

Checking once is not enough ​

Recall the Lancet audit's lead author. He checked his citations, and after several rounds of edits the journal's editor still found a wrong one. STAT does not say how that error got in. In general, though, edits move sentences away from their sources, merge two claims into one, and swap a citation for a near neighbor. A check that runs when you first add a reference does not cover what the draft looks like when you submit it.

What a citation workflow has to guarantee ​

If a lookup covers only the visible end, the useful question about any tool is how much of the chain it keeps. That applies to a reference manager, a writing assistant or an AI agent alike. Here are six tests. You can apply them to anything, including the way you work now with no software at all.

1. Citations come from works you hold. A citation should point to an entry in your own library, with a file or a record behind it, not to a string an assistant produced. If an AI drafts with you, it should cite from that library. A citation to anything outside it should be obvious.
2. A citation says where. It should name a page or a section, not just a work, and that location should exist. Page 312 of a 40-page article is a mistake a machine can catch.
3. The passage is one step away. From the sentence in your draft, you should be able to reach the page it relies on without searching. This is what makes the second failure checkable in practice. You will not check whether a passage supports a claim if finding the passage takes twenty minutes.
4. Reading is on the record. You should be able to see which cited works you have opened, highlighted or annotated, and which you have only cited. That record is not proof that you understood them. It is a record.
5. The check runs at the end. It should run after the last edit, before the file leaves your hands, every time.
6. The report says what it did not check. "No problems found" has to look different from "nothing was checked." "Not in the registry" has to look different from "does not exist."

None of these tests decides whether a passage supports your sentence. That judgment belongs to the person whose name is on the paper. What a tool can do is make the judgment cheap enough to make for every citation, not only for the ones that happen to look suspicious.

How Scholardo answers the six tests ​

Scholardo is a research workbench for macOS and Windows. Your library, your reading and your drafts live in one project folder on your computer, and a built-in team of agents works inside it. Here is how it does against each test, including where the answer is only partial. The export-check guide covers the mechanics in more detail.

1. Citations come from works you hold. A project's bibliography lives in its own references folder. When the Conversation Agent drafts with you, its instructions say to cite with keys from that library ([@citeKey]), to add an entry before citing anything new, and to read a work before citing it. It also leaves the references section empty, and Scholardo fills it from the keys in your citation style. A reference list built that way cannot include a work the library does not hold. Instructions are not guarantees, though. A model can still write a key that does not exist, and that is what the export check is for: a key that does not resolve in the bibliography is reported. When you cite by hand, the selection toolbar's Cite command inserts the key for a source you choose. It does not go looking for a source that supports your sentence.

2. A citation says where. Page locators are checked against the cited file. A page beyond the end of the local PDF is reported. If the page printed in your prose differs from the page recorded as the claim's source, that is reported too.

3. The passage is one step away. When the Conversation Agent writes a sentence that rests on a source, it records an anchor: the sentence, the source and the page. In the check report, the Claims tab lists each anchored claim next to the words on the page it names. One click opens that page. The panel also says, in plain words, that nothing in it judges whether the page bears the claim out. That reading is yours. Here the answer is partial. Anchors are currently recorded by the agent, and there is no way yet to add one by hand for a sentence you wrote yourself. For your own citations, Scholardo checks the locator, but you open the source yourself. The report also tells you how many citations are traced to a page, and warns you when none are.

4. Reading is on the record. Scholardo keeps a record of which documents and pages its agents read through its own tools. It records where and how much, never the words. Two checks use this record. A cited work with no reading record gets a note. An anchor that names page 12, when the record shows page 12 was never opened, gets flagged. The second check is the one that matters for misattribution. It sets the agent's own statement of where a claim came from against a record that Scholardo writes, not the agent. This answer is partial too. The record covers what agents read through Scholardo. Your own reading, in this app, in another one or on paper, is not in it. So a missing record is a question for you, not an accusation.

5. The check runs at the end. It runs on every export unless you turn it off. The Conversation Agent can also run the same check before handing you a draft. If a sentence was rewritten or cut and its anchor stayed behind, the report says the sentence is gone, rather than counting the claim as still traced.

6. The report says what it did not check. The report lists the checks that could not run, so "nothing found" and "nothing examined" do not look the same. Looking a reference up in an external registry is a separate step. You authorize it per project, and it is off by default. When it is on, a DOI lookup can tell you four things: the DOI resolves to nothing, the entry's fields disagree with the registry's record, the DOI points to a different work entirely, or the work has been retracted. A work with no DOI is counted as unchecked, not flagged. That keeps a 1987 monograph from looking like a fabrication.

By default the check reports and lets you export anyway. In blocking mode, each problem needs a written reason before export. All of this is plain files in the project folder: the bibliography, the drafts in Markdown, the reading record, and a per-document evidence record in Markdown. Three months later, or when a reviewer asks where a sentence came from, you open the folder.

Review one paragraph before you submit ​

This is a procedure, not a measured result. Pick the paragraph your argument leans on hardest. Take a sentence like this invented one:

Most respondents reported using shadow libraries every week [@example-survey, p. 14].

1. Run the check on the chapter, either by exporting it or by asking the Conversation Agent to run it.
2. Fix the mechanical findings first. These are unresolved keys, pages past the end of the source, and quotations without a locator. They are quick to fix, and leaving them makes the rest of the report harder to read.
3. Open the Claims tab and read each page next to its claim. Ask whether the source studied the same population, with the same measure, over the same period. Ask whether it is as strong as your wording. "Most" and "a third" are different findings. Check whether page 14 reports the survey's own result or a result it cites from someone else. Citing the second as if it were the first is the improper secondary citation Mogull's review counted, and one of the quietest ways a real paper ends up saying something it doesn't.
4. For citations without an anchor, open the source at the locator yourself. Start with numbers, direct quotations, and the claims your conclusions depend on. Hosseini and a coauthor draw a similar line for hallucinated citations, between those that matter to a paper's conclusions and those that do not.
5. Look at the works with no reading record. If you read one somewhere else, fine. If you can't remember reading it, read it now or cut the citation.
6. If you authorized the registry lookup, read the different-work and retracted findings before anything else.
7. Run the check again after your last edit. Not before it.

What this does not solve ​

A tool that asks you to be honest about evidence should be plain about its own limits.

• It does not judge support. Every result above is about existence, location, coverage and reading. Whether page 12 says what your sentence says is the question from the opening, and Scholardo does not answer it. It puts the page next to the claim. The reading is yours.
• A registry match is not a reading. A DOI that resolves tells you a record exists. It says nothing about the passage, and a lookup that finds nothing is a question, not proof of fabrication.
• The reading record is the agents', not yours. It cannot tell you what you read in Scholardo, in another app or on paper.
• Agents can read your highlights and annotations, but cannot create or change them. What you have marked stays a record of your own reading.
• The check does not review formatting. A valid bibliography does not mean the exported file meets your target journal's requirements.

Back to the question ​

"Does the review panel check the cited passage itself, or only that the citation resolves?"

For the paper you are about to submit, the review panel is you. You are the last person to read it before your name goes on it. Software can tell you whether a citation resolves. It can bring the cited page to the sentence that relies on it, and tell you which sources nobody opened. Reading the passage and deciding whether it holds up is still your job.

A citation is the visible end of an act of reading. Keep the rest of the chain, and a check has something to check. Lose it, and even a reference list where every entry resolves can be wrong in ways no lookup will find.

If you want to see how much of your own chain survives, take a chapter you have already finished and run the review above on one paragraph.

Sources ​

Every quotation in this article comes from the page linked where it first appears. Posts on X were read on November 2, 2026.

• Zhao, Z., Wang, Y., Stuart, T., De Vaan, M., Ginsparg, P., & Yin, Y. (2026). LLM hallucinations in the wild: Large-scale evidence from non-existent citations. arXiv. arXiv:2605.07723
• Topaz, M., et al. (2026). Fabricated citations: An audit across 2·5 million biomedical papers. The Lancet, 407, 1779–1781. doi:10.1016/S0140-6736(26)00603-3
• Mogull, S. A. (2017). Accuracy of cited "facts" in medical research articles: A review of study methodology and recalculation of quotation error rate. PLOS ONE, 12, e0184727. doi:10.1371/journal.pone.0184727
• Pisek, J., & Veromann, E. (2026). AI-hallucinated citations should be treated as research misconduct. Nature, 658, 1356. doi:10.1038/d41586-026-03407-2
• Ithaka S+R. (2026). Buffeted by change: Researchers seek continuity and investment.
• Singh Chawla, D. (2026, May 19). Researchers who use hallucinated references to face arXiv ban. Nature.
• Stokel-Walker, C. (2026, May 14). Hallucinated citations highest in social sciences preprints site. Nature.
• Oza, A. (2026, May 7). Fraudulent citations, blamed on AI hallucinations, are becoming more common in research papers. STAT.
• Retraction Watch. (2026, May 7). One in 277 PubMed-indexed papers in 2026 shows fabricated references, says analysis.
• Van der Sar, E. (2026, October 24). AI-assisted paper on Z-Library cites a TorrentFreak article that doesn't exist. TorrentFreak.

Cover image by Matt Ridley on Unsplash.