Why IEEE retracted 1,400 papers in one day, and how two AI-faked papers got caught

On 23 October 2026, IEEE retracted at least 1,417 conference papers from its IEEE Xplore library in a single day. That is Nature's count; IEEE's own statement of 28 October says "approximately 1,400" and does not list the papers. Research-integrity specialists told Nature it is the largest single bulk retraction by any scholarly publisher. Each retraction notice says the article "may have been offered for sale". The trigger was a study that matched social-media advertisements selling authorship slots, for about US$43 to $400 for a first author, to 2,062 papers in IEEE conference proceedings.

The same week, IEEE Spectrum published Benjamin Skuse's account of faking two research papers with AI and sending them, with the publisher's cooperation, to the journal IEEE Access. Both were stopped. The first was caught by its invented references. The second, written by an autonomous research system, had real references; the journal said it would have rejected it for leaning too heavily on preprints, and its editor spotted a figure with a meaningless label.

Together, the two stories show where integrity checks happen in 2026 and what they can see. Below: what was retracted and why, how the papers were traced, how the AI-made papers were caught, why invented references are a weak point that will not last, what readers can still check, and what a researcher who uses AI honestly can do so that their paper does not look like a mill product. Every source was checked on 2 November 2026.

Why did IEEE retract 1,400 papers in one day?#

Because an outside study matched them to advertisements selling authorship on social media. IEEE's statement, as quoted by Nature, says that "after the irregularities were identified via an external study" it conducted a comprehensive review and found that the papers violated "IEEE's publication principles". According to Research Information, IEEE said it used "new technology and advanced screening methods" and completed the retractions within months of learning of the concerns.

The main numbers, and where each comes from:

The two counts do not conflict: IEEE gives a round figure, and Nature counted the notices. On 2 November we looked up the 2,051 papers in the study's public list that have DOIs in Crossref, the registry where publishers record retractions. Exactly 1,417 carry a retraction dated 23 October 2026, the number Nature counted. In all, 1,515 carry a retraction of some date; Nature, which used the Problematic Paper Screener, counts 1,622.

IEEE did not answer Nature's questions before publication and pointed to its statement. In it, W. Clem Karl, IEEE Vice President, Publication Services and Products, said that "Protecting the integrity of the scholarly record is a responsibility shared across the entire research community". Steven Heffner, IEEE's Chief Publications Officer and Managing Director, Publications, said: "Trust is at the heart of scholarly publishing."

Bulk retractions are not new. In 2015 IEEE retracted more than 1,200 papers and abstracts from a single 2011 conference, Nature notes, and before that IEEE and Springer had pulled some 120 papers made with SCIgen, a program that generates nonsense computer science papers. Across all publishers, retractions passed 10,000 in 2023, and the Retraction Watch Database now lists more than 68,000.

How did researchers trace the papers to paper mills?#

By reading the mills' own advertisements. The study, "Opening Pandora's box: Paper mills in conference proceedings" by Anna Abalkina (Free University of Berlin), Marie Kunešová (University of West Bohemia), Yagmur Ozturk (Univ. Grenoble Alpes) and Solal Pirelli (EPFL), was posted on arXiv on 24 April 2026 and published in Accountability in Research on 1 October.

Starting in July 2025, the authors searched eight social platforms, from Facebook to Telegram and WhatsApp, and the open web for offers of authorship on IEEE conference papers. They collected 5,684 offers, or 4,407 unique titles or keyword sets, from more than 200 channels with a combined audience of over 600,000. A typical offer lists a tentative title, promises Scopus indexing and prices each author slot. In one case, authorships for IEEE conference papers were being sold in the official Facebook group of an IEEE region.

The team then compared the offered titles with the metadata of every IEEE conference paper published since 2021, using edit distance and sentence embeddings, and two co-authors checked every candidate by hand. The result: 2,062 papers in 366 proceedings, 1,913 rated "definitely" and 149 "probably". In 73% of them, the published title matches the advertised one exactly.

What the offers and the papers show:

• Prices. "$43 to $400 for the first authorship, and $11 to $200 for the last," and $217 to $750 for a whole paper.
• Six authors. Offers usually had six slots, and 82.0% of the identified papers have exactly six authors, against 15.6% of all IEEE conference papers in 2024. The authors suspect a reason: IEEE Xplore shows only the first six authors at the top of a paper's page.
• Strangers as co-authors. The papers list 5.1 distinct affiliations on average, and 46% have authors from at least two countries.
• Bought citations. One author was cited 559 times across 112 of the papers, and some references were "unrelated to the topic of the citing paper".
• Fast acceptance. Of 48 conferences with archived websites and at least 10 identified papers, 13 accepted papers within 10 days or fewer.

One conference, IC3I 2022, had 23.51% of its proceedings matched to offers. At the time of the study, only 66 of the 2,062 papers had been retracted.

The authors point to a structural gap. Most of the conferences with five or more identified papers were sponsored by IEEE sections in India, and local organizing committees handle submissions and peer review, Nature reports. Publishers including IEEE receive only "the final “camera-ready” versions of papers", so they are "not aware of potential authorship changes that occur after acceptance". Abalkina told Nature: "Paper mills are a public market and the offers for sale are public on social media. Anyone can find them. Anyone can buy. Anyone can trace [them]."

Were the retracted papers written by AI?#

The study does not say so, and it did not test for it. Its evidence is commercial and bibliographic: advertisements, author counts, affiliations, citations, text reuse, plagiarism and "tortured phrases", odd paraphrases such as "information the executives" for "data management" that often hide copied text. It also describes "generic and superficial content, irrelevant examples, and graphs that do not add meaningful information".

That matters for what comes next. The papers were caught by the traces of a sale, not by anything about who or what wrote the text. A mill can sell authorship on a paper written by hand, by a paraphrasing tool or by a language model, and the same signals apply.

How did IEEE Access catch two AI-faked papers?#

Skuse, a science writer with a PhD in applied mathematics, ran a two-part test with the agreement of Jenny Mahoney, managing editor of IEEE Access, and wrote it up in IEEE Spectrum on 27 October. Neither paper was accepted.

Round 1: free chatbots. Using only free ChatGPT (GPT-5.3/5.2 Mini, he writes) and Gemini 3 Flash, without editing any output himself, he prompted his way to a paper on optical solitons, his old PhD topic, with himself as sole author. "After a dozen or so prompts, I had a realistic-looking paper." The models produced no tortured phrases. The references gave it away. Mahoney ran the paper through STM Integrity Hub, which flagged 12 references, and a second checker from MolecularConnections could not validate 24, one of them by a single author named "John Smith". When Skuse later ran the paper through Gemini, it reported seven fabricated references. "I am an experimental physicist by training, so this was not at all my area, but nothing seemed majorly scientifically wrong," Mahoney said. "It was just that so many hallucinated references immediately cast doubt on the entire work."

Round 2: an autonomous research system. Skuse then used FARS, built by the Singapore startup Analemma, which Spectrum reports was acquired by Baidu while the article was being reported. FARS wrote a research proposal in 21 minutes, ran the computational experiment by the next day and produced a submission-ready paper in two hours. Below the abstract was a line: "WARNING: This paper was generated by an automated research system." Skuse deleted it and put his own name on the paper.

This time the references were real. "While this AI-generated one had no fake references, our team would have rejected this paper because it did have a huge percentage of references to preprints, which likely has something to do with the dataset the AI has access to," Mahoney said. The integrity tools found nothing wrong with the figures, but she did: one had a red line labeled "lhewl", and an AI "brain" in another had random dots in it.

Where do integrity checks happen in 2026?#

At several points, run by different people, each looking at something different. The week's stories cover most of them:

The largest action came last. The papers the study found were published between 2021 and 2026. The study began collecting offers in July 2025, the preprint appeared in April 2026, and the retractions followed on 23 October. That was fast by publishing standards ("It is unbelievable. They're so fast," Abalkina told Nature), but it came after the papers had been published and indexed. In conference proceedings, the authors note, the publisher may never see a paper before its final version.

Submission checks are improving. Mahoney told Spectrum: "It's promising to see the progress that has been made to automate some of the triage tasks, such as author checks, tortured phrases, duplicate submission, reference validity, et cetera." IEEE Access receives about 60,000 submissions a year and accepts about 20%, each with at least two reviewers, and the average number of reviewer invitations per article has risen 50% since 2018. Publishers are now bringing AI to the reviewers' side too: Will AI review my paper? covers the Springer Nature and Google DeepMind assistant and what reviewers may do with AI.

Indexers are a checkpoint too. Retraction Watch reported on 30 October that the US National Library of Medicine has paused MEDLINE indexing and PubMed Central archiving of new Scientific Reports articles while it reevaluates the journal, the largest in PubMed. The journal's website counts 41,122 articles this year; PubMed had indexed 28,593. Neither the chief editor nor NLM had said why. NLM's policy names systemic problems with articles, "verifiable concerns about the scientific or editorial quality" and significant changes such as publication volume as reasons, and says a reevaluation "can typically take around 7 to 10 months". No link to the IEEE case or to AI has been reported.

The rules now name both problems. Among the reasons an editor may lose confidence in a paper, version 3 of COPE's retraction guidelines (August 2025) lists "The authorship of the publication cannot be verified or there are serious concerns about accountability for the research" and "Any form of misrepresentation", including "fraud (eg, a paper mill)" and "undisclosed involvement of artificial intelligence (AI)". For batch retractions it asks notices to "clearly state that systematic, coordinated, and widespread manipulation of the publication process has been identified". It also says that "Some authors could be blameless" and that editors should consider a process of recourse for them, such as re-review.

Luigi Longobardi, IEEE's director of publishing ethics and conduct, gave Spectrum the sober version: "It feels like building a detection tool today may not be effective with the AI-created paper of tomorrow." Ivan Oransky, cofounder of Retraction Watch, gave the vivid one: "publish-or-perish has always been there as kindling, paper mills are the gasoline, and AI is the accelerant."

Why are hallucinated references the weak point of AI-generated papers?#

Because they can be checked against an outside record. A reference can be looked up in Crossref, PubMed or arXiv (with care, since "not found" is not always "fake"); a hedge, a list of three or a formulaic transition cannot. Such stylistic tics, Skuse notes, are "often used by unimaginative humans, too".

The flaw is common. A study by Maxim Topaz and colleagues, published in PNAS on 26 October, sampled 500,000 papers in chemistry, physics, materials science, engineering and computer science from 2023 to March 2026. The share of papers citing at least one title that could not be found in any of seven bibliographic and web sources rose from 1 in 1,600 in 2023 to 1 in 435 in the first quarter of 2026. In May, a team led by Topaz reported in a letter to The Lancet that about 1 in 277 PubMed-indexed papers from the first seven weeks of 2026 cited a paper that does not exist, and that over 98% of them had seen "no publisher action" at the time of the audit.

But it is a flaw of cheap tools, and better tools are losing it. Skuse's second paper, made by a system that searches the literature first, had no invented references. Skuse writes that "more advanced LLMs are no longer making the kind of mistakes their free, dumber relatives are making." Mahoney's objection to that paper, too many preprints, was a judgment about the mix of sources, not a test of whether they exist.

Mohammad Hosseini of Northwestern University told Retraction Watch in May that invented references are "low-hanging fruit". The "bigger and more important problem", he said, is AI-generated citations that are not wholly invented but are inaccurate, biased or incomplete: "We are far from being able to even detect them or do anything about them."

So expect the next generation of fakes to have real references. The checks that will keep working look at what sits outside the text: who the authors are and how they came together, whether data and code exist, how a paper changed between versions, and whether the people named can account for their part.

What can readers and reviewers still check?#

Not whether a machine wrote the text. Here is what still works, drawn from the study and the Spectrum test:

1. Open the sources, not only the links. A reference that resolves can still say something else. Our guide on how to check for hallucinated references has a step-by-step list.
2. Look up retraction status. Crossref's record for a DOI shows retractions the publisher has registered, and the Retraction Watch Database lists more than 68,000. The study also used the Problematic Paper Screener, which flags tortured phrases and citations of retracted work.
3. Read the author list. Six authors from five unrelated institutions proves nothing on its own. Of the rise in six-author papers, the study writes: "Though this pattern is not a sign of misbehavior, it may indicate that paper mills have infiltrated many conferences." An author list that no shared project explains is worth a question.
4. Look at the venue. Proceedings range from the most competitive in computing, such as CVPR and ICASSP, to conferences that accepted papers within days. The acceptance timeline and who ran peer review are often on the conference website.
5. Look at the figures. Labels like "lhewl", axes without units and illustrations that add nothing were among the tells in both stories.
6. Find the AI disclosure. IEEE requires that AI-generated content "shall be disclosed in the acknowledgments section", naming the system and the sections. A disclosure tells you what the authors say they did. Its absence proves nothing either way.
7. Ask for the data. The study published its offers, matched papers and scripts on Zenodo, which is how we could rerun part of its count in Crossref.

How do you keep an AI-assisted paper from looking like a mill product?#

Publishers such as IEEE allow AI use when it is disclosed; hiding it, or selling authorship, is grounds for retraction under COPE's guidelines. A researcher who works with AI and wants their paper to stand up to the checks above can do five things:

• Disclose, in the place the venue asks. At IEEE that is the acknowledgments: the AI system, the sections, and "a brief explanation regarding the level at which the AI system was used". The study behind the retractions shows the form in two sentences: "The authors used Gen AI for grammar checking only, not for editing or writing the text. The authors used OCR technology, including Gen AI models, to identify the text in images, and the results were manually checked." Our guide to AI disclosure statements has more examples, and Journal AI policies quotes what 29 publishers, journals, preprint servers and guidelines ask.
• Verify every reference yourself. Do not let an AI supply references from memory, and check the ones it only reformatted.
• Keep your versions. Drafts, data, code, prompts and outputs. Elsevier's policy, as summarized on our journals page, already asks authors to keep a separate record, which may include logs, saved prompts and outputs.
• Make the author list match the work. Every author should be able to say what they did. Be wary of names added after acceptance: in conference proceedings, the publisher may only see the final version.
• Never remove a tool's own disclosure. Skuse deleted FARS's warning line to test the system. In a real submission, sending the paper without saying how it was made would be the "undisclosed involvement of artificial intelligence (AI)" that COPE lists as misrepresentation.

If a paper is caught up in a batch retraction, COPE's guidelines say editors should consider recourse for blameless authors. Drafts, data and a dated account of who did what are what you would bring to that conversation.

Where does Papex fit?#

Papex keeps a record of how research is made. AI apps can prepare work for it; you review it, sign it and decide what to publish. An AI app connected to Papex, such as Claude or ChatGPT, can prepare a private draft, propose changes, look up references by DOI, arXiv ID, PMID or ISBN, and leave a private summary of each working session that only you see. It cannot sign, accept its own proposals or publish, and it is never listed as an author. Model and tool names are kept as the app reports them and labeled Agent-reported.

When you sign, the signature covers one exact version, and a salted fingerprint of that version is timestamped with OpenTimestamps and anchored in Bitcoin. Anyone can check the signer, version and date at /verify without an account. From the signed version, Papex can write the AI-use statement in a journal's own form: IEEE, ACM, arXiv and others are listed on /docs/journals.

Papex does not detect paper mills or AI-written text. A signed record shows who signed which version, when, and what they said the AI did; it is not peer review and does not show the work is correct.

What comes next?#

The study asks conference publishers to screen papers across their whole life, not only the final version; to monitor social media for authorship sellers; to speed up investigations of organized manipulation; and to consider retracting whole compromised conferences. IEEE's statement, as summarized by Knowledgespeak, lists wider screening, work with the STM Integrity Hub and COPE, and volunteer experts in investigations.

Detection will keep improving, and so will the fakes. In Spectrum, Longobardi argues that AI may ultimately require scientists to prove not just what they discovered but how they discovered it. The researchers best placed for that are those who disclosed their AI use, kept their versions and can say who did what. That is the record Papex is built to help keep.

Sources#

• Miryam Naddaf, Record-breaking retraction: scholarly publisher pulls 1,400 conference papers on one day (Nature, 30 October 2026), checked 2 November 2026 (the part after "Conferences targeted" is behind Nature's login and was not read)
• IEEE, statement on the retractions (IEEE, 28 October 2026), checked 2 November 2026; the page blocked automated reading and has no Internet Archive copy, so its wording is taken from Nature, Research Information and Knowledgespeak
• IEEE retracts approximately 1,400 papers following integrity review (Research Information, 30 October 2026), checked 2 November 2026
• IEEE retracts 1,400 Papers from IEEE Xplore following research integrity review (Knowledgespeak, 29 October 2026), checked 2 November 2026
• Anna Abalkina, Marie Kunešová, Yagmur Ozturk and Solal Pirelli, Opening Pandora's box: Paper mills in conference proceedings (Accountability in Research, 1 October 2026), checked 2 November 2026; the journal's site blocked automated reading, so we read the arXiv version 2 of 11 August 2026
• Abalkina, Kunešová, Ozturk and Pirelli, Data for "Opening Pandora's box" (Zenodo, 10 July 2026), checked 2 November 2026
• Crossref REST API, retraction metadata for the study's 2,051 DOIs, queried 2 November 2026
• Benjamin Skuse, I Used AI to Fake a Scientific Paper. Here's What Happened (IEEE Spectrum, 27 October 2026), checked 2 November 2026
• Kate Travis, Scientific Reports on hold in PubMed while NLM reevaluates journal (Retraction Watch, 30 October 2026), checked 2 November 2026
• National Library of Medicine, MEDLINE Policies: Reevaluation (NLM), checked 2 November 2026
• Retraction Watch Staff, Weekend reads: Mega-journal on hold at PubMed; 1,400 retractions in one day; 'I Used AI to Fake a Scientific Paper' (Retraction Watch, 31 October 2026), checked 2 November 2026
• COPE Council, COPE Retraction guidelines, version 3 (COPE, August 2025), checked 2 November 2026; the live site blocked automated reading, so we read the Internet Archive copy of the PDF dated 6 July 2026
• Rita Aksenfeld, New COPE retraction guidelines address paper mills, third parties, and more (Retraction Watch, 4 September 2025), checked 2 November 2026
• IEEE Author Center, Submission and Peer Review Policies: Guidelines for Artificial Intelligence (AI)-Generated Content (IEEE), checked 2 November 2026
• Maxim Topaz, Pallavi Gupta, Zhihong Zhang and Laura-Maria Peltonen, Evaluation of possibly fabricated references in 500,000 science, engineering, and computing papers (PNAS, 26 October 2026), checked 2 November 2026; the journal's site blocked automated reading, so we read the abstract in its Crossref record
• Avery Orrall, One in 277 PubMed-indexed papers in 2026 shows fabricated references, says analysis (Retraction Watch, 7 May 2026), checked 2 November 2026
• Alison McCook, One publisher appears to have retracted thousands of meeting abstracts. Yes, thousands. (Retraction Watch, 25 June 2015), checked 2 November 2026
• Richard Van Noorden, More than 10,000 research papers were retracted in 2023 — a new record (Nature, 12 December 2023), checked 2 November 2026