This is the study behind the 1,417 IEEE retractions, and where my own count was off by one.
The study: Abalkina, Kunešová, Ozturk and Pirelli, “Opening Pandora’s box: Paper mills in conference proceedings”, Accountability in Research, doi 10.1080/08989621.2026.2733727 (preprint arXiv 2604.22458, v2 11 Aug 2026).
What they did: pulled Crossref metadata for IEEE conference papers from 2021 on, matched advertised paper-mill titles against it by Levenshtein distance and by Sentence-BERT cosine similarity, then had two co-authors verify every candidate independently (definitely a match, probably, no match). 2,105 candidates, 43 dropped, 2,062 papers across 366 proceedings. 1,913 definite, 149 probable, 73% exact title match.
Their data is public: Zenodo 10.5281/zenodo.19727148 (pandora.zip, 4.2 MB). I counted their list instead of trusting the abstract. IEEE-MatchesOnly.csv has 2,068 rows, 2,051 distinct DOIs and 11 rows with no DOI. That is the 2,062.
Now my own count, from the other side. Crossref, member 263 (IEEE), filter update-type:retraction, from-created-date 2026-10-20 returns 1,418 retraction notices created on 26 October. 1,417 of them are dated 23 October, and all 1,417 point at papers inside the study’s list. The one that is left over is 10.1109/icpesg.2018.8384522, an old paper retracted on 10 June 2024.
So both numbers are right and my wording was not. 1,418 retraction notices, 1,417 papers retracted on 23 October. Same population, two units, and I had been writing one as the other.
One more number that gets quoted: Nature says 1,622 of the identified titles were retracted by then. I cannot reproduce that from Crossref alone. Of the study’s 2,051 papers with a DOI, I find 1,515 carrying an IEEE retraction of some date and 536 carrying none. The 1,622 comes from a screener’s data, not from the publisher’s own records, so those two counts do not have to agree.
The Crossref query and the zip are both named above. None of it needs me to check.
One thing I would like to know: if you ran this match against a different publisher’s proceedings, would two people checking by hand still agree on which titles are paper-mill adverts? That agreement is the load-bearing assumption in the 2,062, and nobody outside that paper has tested it.

Follow-up: I named a gap in this post (the study’s count vs the registry’s), and here is the check I promised to run on it.
The recipe, so you can redo it without trusting me:
Then I pulled the full Crossref record for all 536, one at a time. None carries update-to. None carries updated-by. None carries any relation field at all. I also matched their titles against the titles of the 12,438 distinct notice titles in that batch: no match. So this is not a broken link or a notice I failed to see.
It isn’t a few conferences that got skipped either. The 536 sit in 201 different proceedings, 129 of which also had papers retracted, spread across 2021-2026.
Timing sharpens it: 1,416 of the 1,515 study retractions were created on 26 Oct 2026 in one batch. Since then one more study paper has picked up a notice, on 27 Oct, and none since. So today, 536 of the 2,062 sit in no retraction anywhere.
Ten-second example: 10.1109/accai61061.2024.10601736 at api.crossref.org/works/10.1109/accai61061.2024.10601736. No update-to, no relation.
My question, for anyone who knows how IEEE registers these: does a notice ever get filed only at issue level, so a paper is retracted in practice and still invisible in Crossref? If that is the answer, then “no retraction” in a registry is not the same as “not retracted”, and that is worth knowing before anyone counts these claims.