Twenty of Twenty-One DOIs Resolve. The Trouble Is in the Other Twenty-Three References.

On 24 October TorrentFreak reported that a paper in the New Review of Academic Librarianship, a Taylor & Francis journal, cites a 2021 TorrentFreak article called “Publishers continue crackdown on shadow libraries.” TorrentFreak says it never published an article with that title. The paper is “Z-Library and Full-Text Access: Democratizing Knowledge or Undermining Copyright Ethics?”, by two university librarians in Nigeria, published on 6 October. Its acknowledgements say “Consensus AI was used in sifting through large volume of literature,” and that ChatGPT helped with the language. TorrentFreak could not locate at least 10 of its 44 references. Taylor & Francis told TorrentFreak the concerns had gone to its publishing ethics team.

A count of 10 out of 44 says how many references failed but not where. I pulled the reference list from Crossref, where the publisher deposits it, and checked each entry myself.

The references with DOIs

Twenty-one of the 44 references carry a DOI. I resolved every one. Nineteen return a Crossref record whose title matches the reference, and one more, a Zenodo report on diamond open access journals, resolves through DataCite. One, 10.1007/s11192-020-03570-6, returns a 404 at doi.org. So 20 of the 21 DOI-bearing references point at a real work.

Crossref marks all of these DOIs as asserted by the publisher, not by the authors. My reading is that the publisher’s reference matching attached a DOI wherever it found the cited work, so a reference with a DOI is a reference that matching succeeded on. That makes the 20 resolutions less reassuring than they look. They are the references that had already passed one check.

The 44 references of the Z-Library paper, by whether they carry a DOI and whether it resolves One bar of 44 references split into three segments: 20 with a DOI that resolves to a real work, 1 with a DOI that returns 404, and 23 with no DOI. Under the no-DOI segment, labels note that TorrentFreak could not locate at least 10 references, and that 2 resolve in Crossref to real papers cited with the wrong title or the wrong year and volume. 44 REFERENCES · CROSSREF RECORD, CHECKED 24 OCT 2026 20 references with a DOI that resolves — Crossref and DataCite, 24 Oct 2026 1 reference with a DOI that returns 404 — doi.org, 24 Oct 2026 23 references with no DOI — Crossref, 24 Oct 2026 20 with a DOI that resolves 23 with no DOI 1 dead DOI At least 10 could not be located (TorrentFreak) 2 are real papers cited with the wrong title, or the wrong year and volume (Crossref) I did not check whether each reference supports the claim it is cited for; TorrentFreak says several do not.
Almost everything that carries a DOI is real. The fabricated and garbled references are concentrated in the 23 that do not, which is also where no automated link check would have looked.

Half-right is harder to catch than invented

The other 23 references have no DOI. TorrentFreak’s examples come from this group: a Harvard Journal of Law & Technology article credited to Daniel Gervais that is not in the cited issue, a College & Research Libraries article credited to Carol Tenopir that is not in the cited issue, the TorrentFreak article, and a Justice Department press release under a title the department did not use.

Two more in this group did not look wrong to me until I searched Crossref for them. The paper cites Greshake (2017) as “The Sci-Hub effect: Evidence of downloading and usage impact on scholarly publishing,” F1000Research 6, 541. Bastian Greshake did publish in F1000Research volume 6 at article 541 in April 2017, but under the title “Looking into Pandora’s Box: The Content of Sci-Hub and its Usage.” The paper also cites Lewis (2018), “The inevitability of open access,” College & Research Libraries 79(3), 255–270. David Lewis’s article of that title exists, but in volume 73, issue 5, pages 493–506, in 2012.

Two references that point at real papers with the wrong details A two-row ledger comparing each citation as printed in the paper with the Crossref record. Greshake 2017: cited as "The Sci-Hub effect: Evidence of downloading and usage impact on scholarly publishing", F1000Research 6, 541; the record at that volume and article number is "Looking into Pandora's Box: The Content of Sci-Hub and its Usage". Lewis: cited as College and Research Libraries 79(3), 255 to 270, 2018; the record is volume 73, issue 5, pages 493 to 506, 2012, same title. AS CITED IN THE PAPER VS THE CROSSREF RECORD As cited Crossref record Greshake (2017) Lewis "The Sci-Hub effect: Evidence of downloading and usage impact on scholarly publishing" F1000Research 6, 541 "Looking into Pandora's Box: The Content of Sci-Hub and its Usage" F1000Research 6, 541 · 21 April 2017 "The inevitability of open access" College & Research Libraries 79(3), 255–270, 2018 "The Inevitability of Open Access" College & Research Libraries 73(5), 493–506, 2012 Same author and journal in both rows. Each citation gets one field wrong.
Both references survive a "does this author have a paper in this journal" check. Each fails on one field: the title in one case, the year, volume and pages in the other.

These two are a different failure from an invented source. The work is real, so a reviewer who recognises the author and the journal sees nothing odd, and a search on the author’s name finds a matching paper. Only the title, or the volume and year, is wrong. Published research has measured both kinds. Walters and Wilder, in Scientific Reports in 2023, had ChatGPT write short literature reviews on 42 topics and checked all 636 citations: 55% of GPT-3.5’s citations and 18% of GPT-4’s were fabricated, and of the real ones, 43% and 24% respectively had substantive errors.

Fabricated citations and errors in real citations, GPT-3.5 and GPT-4, Walters and Wilder 2023 Horizontal bars on a 0 to 100 percent scale from 636 citations in 84 short literature reviews. Fabricated citations: GPT-3.5 55 percent, GPT-4 18 percent. Substantive errors among the real citations: GPT-3.5 43 percent, GPT-4 24 percent. 636 CITATIONS · WALTERS & WILDER, SCIENTIFIC REPORTS, 2023 Citations that were fabricated Real citations with substantive errors GPT-3.5 GPT-4 GPT-3.5 GPT-4 55% — Walters & Wilder, Scientific Reports, 7 Sep 2023 18% — Walters & Wilder, Scientific Reports, 7 Sep 2023 43% — Walters & Wilder, Scientific Reports, 7 Sep 2023 24% — Walters & Wilder, Scientific Reports, 7 Sep 2023 55% 18% 43% 24% Short literature reviews on 42 topics. Error rates are among citations to real works.
The newer model invented far fewer sources, but about a quarter of the real sources it cited still had a substantive error, the same kind of fault as the Greshake and Lewis entries.

What a journal could check

None of this tells me how the references got into the paper, and the authors did not answer TorrentFreak’s follow-up questions. The check itself is cheap, though. Pulling a deposited reference list and resolving its DOIs took a few minutes, and it cleared 20 of the 44 references. The other 23 are where a person has to look, because they are free text that no resolver touches, and every fabricated or garbled reference I know of in this paper is among them.

The Lewis entry is a good test of any such process. Its title is right, its author is right, and its journal is right. A check that stops at “does this paper exist” passes it. A check that asks whether the paper appeared in that volume in that year catches it.

References

  1. Van der Sar, E. (2026). AI-Assisted Paper on Z-Library Cites a TorrentFreak Article That Doesn’t Exist. TorrentFreak, 24 October 2026.
  2. Aregbesola, A. & Owolabi, S. E. (2026). Z-Library and Full-Text Access: Democratizing Knowledge or Undermining Copyright Ethics?. New Review of Academic Librarianship, published 6 October 2026. Reference list read via Crossref, 24 October 2026.
  3. Greshake, B. (2017). Looking into Pandora’s Box: The Content of Sci-Hub and its Usage. F1000Research 6, 541, 21 April 2017.
  4. Lewis, D. W. (2012). The Inevitability of Open Access. College & Research Libraries 73(5), 493–506.
  5. Walters, W. H. & Wilder, E. I. (2023). Fabrication and errors in the bibliographic citations generated by ChatGPT. Scientific Reports 13, 14045, 7 September 2023.