OpenNeuro Scans Become Camouflage for Fake Brain Papers

Dorothy Bishop recognized her own OpenNeuro fMRI dataset in a manuscript she was asked to review in August 2026. The scans came from 50 people doing language tasks, a pile of raw data that takes serious computing to analyze.

That journal rejected the manuscript. A name search then found two 2026 papers in Elsevier’s Journal of Neurolinguistics that claim to analyze OpenNeuro brain data Bishop and a second data owner say was never properly used.

A Reviewer Recognized Her Own fMRI Dataset

Bishop is an emeritus professor of developmental neuropsychology at the University of Oxford. She had posted the scans four years earlier, and she was curious that anyone had already finished the work. She told interviewers the description in the manuscript did not match what her group had done, and several references did not support the claims they were cited for.

The editor was suspicious too, she said, and the paper was rejected. Bishop still expected the authors to try another journal, so she searched their names. The trail led to Afzal Khan, who teaches in the English Language Skills Department at King Saud University in Riyadh, and to Dawood Khan of Iqra University in Pakistan.

HOW THE CASE UNFOLDED

1. March 16, 2022: Bishop’s team deposits its language laterality fMRI set on OpenNeuro as ds004073.
2. September 15, 2023: Joe Bathelt, Joanne Taylor, and Kathy Rastle deposit ds004765, a resting-state language fMRI set.
3. August 2026: Bishop is asked to review a manuscript that uses her set; she also finds the first Khan paper in the journal’s volume 79.
4. September 2026: Bishop posts on PubPeer, contacts the authors of a children’s language set, and asks the journal to retract.

On her blog she wrote that a Python script attached to the rejected manuscript did not explain the processing and would not even run on dummy code. The text still looked plausible. The question was original, the results were messy rather than too neat, and she wondered whether she would have caught it if the scans had not been hers.

This is clearly some sort of AI that knows enough about a dataset to produce a plausible paper that purports to use it.

Dorothy Bishop, emeritus professor of developmental neuropsychology, University of Oxford

Kathy Rastle, a psychologist at University of London, looked at the paper that named her group’s data and reached the same conclusion. “I was flabbergasted that someone could take the time to generate a paper that I believe is fake,” she said.

Two Papers Rest on OpenNeuro DOIs They Misread

The first article, “Decoding cognitive diversity in reading,” sits in volume 79 as article 101346, dated August 2026. Khan and Dawood Khan say they used a graph autoencoder on resting-state scans from 69 neurologically typical adults aged 18 to 40, and they report three latent subtypes of reading-related brain networks. They cite OpenNeuro ds004765 and thank Bathelt, Taylor, and Rastle by name.

Rastle’s group did share a language fMRI dataset on OpenNeuro under that accession number. The public README describes resting-state scans and reading-related behaviour in 69 adults; the repository lists 71 participants. That near-match is not the problem she flagged. The published paper said behavioural data were available for only some people, when they were available for everyone. It also reported reading comprehension scores the dataset does not contain. “We’re not sure how they generated the statistics for these variables that didn’t even exist,” Rastle said. A figure of the brain regions involved in reading, she added, “just bore no resemblance to what we know about reading in the brain.”

WHAT THE PAPERS CLAIM VERSUS WHAT THE ARCHIVES HOLD

Rastle does not think the authors used her scans at all. She thinks they named the set “in order to create a veneer of respectability.” Bishop’s view of both papers is that the authors may have downloaded a sliver of data to make a figure, and not much more.

Khan Confirms the Children’s Scans Were Timed Wrong

The second article, under Khan’s name alone, is the connectome paper still listed by Elsevier as article 101364 in the November 2026 issue. It claims connectome-based predictive modelling and representational similarity analysis on 477 people: 48 monolingual adults, 52 bilingual adults, 321 typically developing children, and 56 German-speaking children with dyslexia. It reports that bilingual connectivity predicted reading at r = 0.429 (p = 0.001), that children’s connectivity predicted reading at r = 0.117 (p = 0.018), and that monolingual and dyslexic models failed.

Bishop raised comments posted on the November paper and said Khan “had a total fiasco attempting to respond.” Sample sizes did not match the OpenNeuro sets. Khan wrote that he had dropped people with incomplete scans, including Language History Questionnaire cases 3003 and 3020, but Bishop replied that the transparency statement still said no one was excluded after preprocessing, and that the counts still failed to line up.

ERRORS THAT DO NOT LOOK LIKE TYPOS

• Impossible L2 years: One graph shows a person using a second language for more than 50 years, against a maximum of 37 in the data Khan said he analyzed. He insisted the graph does not go above 50 and wrote, “The reported results remain valid as published.”
• Wrong scanner timing: Bishop asked Jin Wang, an author of OpenNeuro ds003604, a longitudinal language study of children aged 5, 7, and 9. Wang found that Khan’s pipeline used a 2.5 second repetition time. The dataset’s own JSON files give 1.25 seconds.
• A result that depended on that error: Khan later called Wang’s catch “decisive” and said it affects filtering and connectivity for the whole children’s group, “which underlies the paper’s only unambiguously significant, non-bilingual CPM result.”
• A résumé that jumped fields: Khan’s own author note lists a linguistics PhD and more than 19 years in higher education. Bishop said his record before 2026 was largely English teaching and literary criticism, with no sign he had the computing background these scans require.

This was flagged by the dataset’s own author, not something I had identified independently beforehand, and I want that distinction to be clear.

Afzal Khan, author response on PubPeer, September 2026

He told Bishop and Wang he would redo some analyses and that a correction was being processed. In email he said he was “taking them seriously” and “engaging directly and in good faith with the journal,” and that he could not discuss details while the review ran.

The Public Archive That Invites Secondary Analysis

OpenNeuro is a BRAIN Initiative archive run by the Stanford Center for Reproducible Neuroscience. It accepts MRI, PET, MEG, EEG, and iEEG data only in BIDS form, mints a DOI for each snapshot, and, after a grace period, releases sets under a Creative Commons CC0 license after upload. Journals and funders point authors there so other groups can rerun or extend the work.

OPENNEURO AT THE TIME OF THE PAPERS

• Public datasets: 1,909 sets are listed as public.
• People in those sets: 90,449 participants across modalities.
• MRI share: 1,270 public MRI datasets covering 57,319 people.
• Reuse design: Each version gets a DataCite DOI so a secondary paper can cite the exact snapshot.

That design is the point of the archive, and it is also the opening. A public page carries the N, the tasks, the authors, and a how-to-acknowledge line. A model that has seen those pages can write a paper that names the right dataset, thanks the right people, and reports numbers that sit close to the real ones without ever running the pipeline. Bishop’s check for the rejected manuscript, data plus code plus references plus common sense, is the same kit she now says is falling behind. “The frauds are getting so good that these methods now seem inadequate,” she wrote.

The cost does not fall on the mill first. It falls on the lab that spent years collecting the scans and on the people who lay in the magnet. Those labs are now doing the forensic work journals did not do at submission, and they are doing it on data they already gave away.

Formula Papers Already Swarm Open Health Data

Brain images are heavy files, and they are supposed to be a poor target for industrial-scale reuse. Spreadsheet archives are not. Matt Spick, a statistician at the University of Surrey, and colleagues documented a surge of formulaic NHANES association papers that take one predictor and one health outcome from the U.S. National Health and Nutrition Examination Survey and write it up as a finding. Their PLOS Biology search found 341 such papers in 147 journals. From 2014 through 2021 the average was 4 a year. The count was 33 in 2022, 82 in 2023, and 190 in 2024 through 9 October.

A later scientometric follow-up in the Journal of Clinical Epidemiology, again with Spick among the authors, flagged nine public health and biomedical resources with the same hallmarks, including NHANES, UK Biobank, FinnGen, and the Global Burden of Disease Study. Those nine sources had 23,005 papers indexed in OpenAlex in 2025, a 3.0-times rise from 7,655 in 2022, and 11,577 papers above an ARIMA forecast of the earlier trend. PLOS later told editors to reject submissions that rest only on such databases unless the authors add new confirmatory work.

The neuroimaging case is smaller and slower, because an fMRI set is not a CSV file. The trick is the same. Cite a real, named, DOI-stamped public resource, write methods that sound like the field, and let the archive’s reputation travel with the paper. Health journals are already slamming that door. Language-imaging journals have not.

Elsevier Has Opened a Review That May Last Months

Bishop asked the editor of the Journal of Neurolinguistics to retract. Elsevier told her it would investigate and that the process “may take considerable time.” A spokesperson later confirmed both papers are under review. Journal editor Nai Ding said the journal had sent its own inquiry to Elsevier for a further look.

Rastle’s reply was that a paper this flawed should not sit in the literature for months. “It’s just polluting the scientific record and creating confusion amongst readers and other scientists,” she said. The articles remained on ScienceDirect when checked, with no retraction notice on either record.

WHAT WE KNOW

• The record: Two Journal of Neurolinguistics articles in volumes 79 and 80 claim OpenNeuro analyses and remain live.
• The owners: Rastle says her group’s variables were misstated; Wang says the children’s TR was doubled; Bishop says a third manuscript using her own set did not run.
• The author: Khan has accepted the TR error, promised a correction, and declined further comment during the publisher’s review.

WHAT IS UNCONFIRMED

• How the text was made: Bishop and Rastle believe a model wrote the papers. That is their reading of the errors, not a finding by the journal.
• What Elsevier will do: Retraction, correction, or no action has not been announced, and the publisher has already said the work may run a long time.
• How many similar papers exist: Bishop found these by searching two names after one review. No one has audited OpenNeuro citations at NHANES scale.

Khan is still listed on both articles as a researcher at the intersection of neurolinguistics and connectome modelling. The author notes on the papers now read that way. His earlier output, Bishop said, does not.

Participants Shared Scans to Help Science, Not Fraud

Open data in this field was a hard sell. People in the scanner signed up so other scientists could test new questions on the same brains. Funders then made sharing a condition of the grant. The archive did its job: the DOIs work, the BIDS folders validate, and anyone can download ds004765 or ds003604 without asking the original lab.

That success is now the vulnerability. A reuse paper that actually ran the data would still be welcome. A reuse paper that only borrows the names of the people who collected it is a tax on those people. They have to reread their own README, check a TR in a JSON file, and write PubPeer comments that the publisher may sit on for months. Bishop called the episode “very upsetting for those of us who are into open data.”

Participants allow their data to be shared openly to further scientific progress. They don’t allow it to be shared openly to perpetuate fraud.

Kathy Rastle, psychologist, University of London

The two papers are still in the journal. The scans they cite are still public, which is what the volunteers were told would happen. The unpaid work of showing that the analyses do not match those scans has already been done by the labs that collected them.