Bring Sanity to the scientific record.
Our mission: we study how retracted and unreliable research keeps spreading through the published literature, and publish every result, dataset and method free to the public. SaneCite sells nothing, and once it has a nonprofit fiscal sponsor every dollar in and out is public. Read the commitments →
63.5% of retracted papers are cited again after they’re pulled.
32,701 of 51,474 traced · Campaign One · count of 21 Oct 2026
Retracted papers keep getting cited.
Hardly anyone reads a retraction notice. In one large study of biomedicine, only 5.4% of sentences citing a paper after its retraction said it had been retracted.1
of retracted papers were cited again after the retraction: 32,701 of 51,474.
citations, in all, to papers that had already been retracted.
retracted papers were still being cited in 2025 or 2026.
Source: Retraction Watch database via Crossref, cross-checked with the OpenAlex citation graph. Updated 21 Oct 2026, 10:45 UTC. What these numbers don’t say →
Retracted2 Feb 2010
Ileal-lymphoid-nodular hyperplasia, non-specific colitis, and pervasive developmental disorder in children
The paper behind the MMR vaccine scare.
OpenAlex yearly counts start in 2012. *2026 is year to date.
Many of these cite the paper to warn about it. Some may still cite it as support. A count can’t tell which, so Phase 2 reads every one.
Show the numbers
| Year | Citations |
|---|
Fake papers are growing faster than real science.
Researchers at Northwestern University traced organised networks that sell fake papers and authorships, and found them growing faster than legitimate publishing (PNAS, August 2025). Catching a fake paper is half the job. SaneCite measures the other half: whether science keeps building on it after it is retracted.
Then three AI models read 55,688 citations to retracted biomedical papers.
Every citation made after retraction to a PubMed-indexed paper, wherever the citing paper is open access. All three models read each citing passage; on a blind sample they matched a separate AI reader 94.7% of the time. These figures are pending human review.
40.7% of citations to retracted papers still treat them as support.
Out of every 100 citations made after a paper was retracted, 41 treat it as support, 56 mention it in passing, and just 3 flag it as retracted or unreliable.
Out of every 100 citations:
- 41treated it as support40.7% · 21,037 of 51,626
- 56mentioned it in passing55.9% · 28,863 of 51,626
- 3flagged it as retracted or unreliable3.3% · 1,726 of 51,626
Source: SaneCite Phase 2 census: 51,626 citations to 12,328 retracted papers, from open-access papers citing retracted biomedical papers (PubMed-indexed, retracted 1990–2023). Read by three AI models; pending human review. Data as of 14 Oct 2026.
Source: SaneCite Phase 2 census, 51,626 citations, data as of 14 Oct 2026 · sanecite.com/charts
Show the numbers
| Label | Citations | Share (95% CI, %) |
|---|---|---|
| Treated it as support | 21,037 | 40.7% (39.9–41.5) |
| Mentioned it in passing | 28,863 | 55.9% (55.1–56.8) |
| Flagged it as retracted or unreliable | 1,726 | 3.3% (2.6–4.3) |
| All judged | 51,626 | 100% |
51,626 citations to 12,328 retracted papers, each made in a year after the retraction. Left out: 4,044 citations to paperwork retractions (no fault found in the science) and 18 the three readers could not decide. Not corrected for reader error. Corrected using one AI reader's checks of 150 citations, the estimate is 38.1% (95% CI 33.5–41.8%). 95% confidence intervals (the likely range) come from re-running the numbers 400 times, drawing whole retracted papers at random (a cluster bootstrap).
Even when a paper’s published reference list showed the retraction, 2 in 5 citations still treated it as support (40.4% of 3,628).
The 40.7% and 3.3% are shares of the 51,626 citations the panel could judge: the 55,688 read, minus citations to paperwork retractions (out of date, replaced, moved journal; reported apart) and 18 it could not call.
Citations made more than ten years after a retraction still treat the paper as support 34.9% of the time (43.6% in the first two years; the mix of papers differs by year). Papers retracted for misconduct were cited as support about as often as those retracted for error. Method, limits and data → · Search every citation → · All charts, free to share →
Ordinary citations go wrong too. About 1 in 14 badly misreport their source.
We drew 1,000 random citations from open-access papers in PubMed Central and read each one next to the paper it cites, with the reading rules and pass marks fixed before the test ran.
of the 915 citations that could be judged badly misreport the paper they cite (66 of 915; 95% CI 5.7–9.1%).
have some error, from an overstated finding to a source that says the opposite (186 of 915).
of the serious errors were caught by the checker, at $11–13 per 1,000 citations. That is what makes reading every citation affordable.
Citations that credit a specific finding are the riskiest in two separate samples: 9.7% and 7.8% of them badly misreport it. These figures are measured against AI readers, not people. How the test worked →
Two things stand in the way.
A machine reads it all. You can check the machine.
1Cheap to verify.
Every call is published with its passage, so anyone can check it. The money goes to people who check the machine.
| Experts checking the machine’s flagswhere most of the money goes | most of it |
| Campaign One, all three phases | under $50,000 |
Already run on real papers. We have read and checked 68,022 citations so far: 63,026 citations of retracted papers, including the 55,688 in the published Phase 2 census, and 4,996 ordinary citations in our accuracy tests. Every census call is published with its passage, for anyone to check.
2Open, so you don’t have to trust us.
A citation is a claim about a source. SaneCite reads the source, and every verdict quotes the passage it rests on.
The method is published before each run, and the raw results come out with the findings. The code behind the Campaign One census is open source, so anyone can re-run it.
The census code (the Phase 1 count, the Phase 2 reading panel and the model calls) is public under the MIT license, with its data, on GitHub. The code of the checker used in the accuracy test is not public yet.
The retraction audit, in three phases.
Every retracted paper is in the set because it was retracted, not because we picked it. Nobody gets singled out.
- Done
1Count
Match every retracted paper to every citation it got after the retraction. Mechanical, no judgement calls.
Out: the numbers above. Tables by journal and field, and the full dataset, publish next.
- Census read · pending human review
2Read
For every citing paper with open full text, read the passage that cites the retracted work. Did the author use it as support, mention it in passing, or flag it?
Census: 40.7% of 51,626 citations cited the retracted paper as support.
- After Phase 2
3Tell
Send each journal that published a citation resting on retracted work the evidence. Then publish who corrected, who replied, and who said nothing.
Out: a public ledger of replies.
What a count does not tell you.
- A flag is not a finding.Some papers cite a retracted paper to criticise it. Phase 1 can’t see the difference. The Phase 2 census can: 40.7% of citations treated it as support, and most of the rest mentioned it in passing.
- Authors often couldn’t know.The notice never reached them. We measure that failure, of the system, not of the scientists.
- Paperwork retractions are left out.Some retractions find no fault in the science: a review withdrawn as out of date, a temporary removal, a publisher’s error, a paper replaced at the same address. Since 16 Oct 2026 the count leaves out such papers (1,856 papers and 23,520 citations on 21 Oct 2026), the same rule the Phase 2 census always used. Updated 16 Oct 2026: the first count, 286,050, included them.
- We count by year, so we undercount.Only citations in the years after the retraction year are counted. Citations in the same year are left out.
Coverage: 53,373 of the 53,732 retracted papers that carry a DOI (the permanent ID number a paper gets when it is published), retracted in 2024 or earlier, matched in OpenAlex (21 Oct 2026); 51,474 after leaving out paperwork retractions. Retraction Watch holds 67,891 retraction records in all (21 Oct 2026). Not every counted citation is real: OpenAlex sometimes links the wrong paper, counts a retraction notice, or lists one paper twice. We checked 200 counted citations at random against the citing papers’ own reference lists (16 Oct 2026): about 96% were real, so the true total is about 254,000 (likely between 247,000 and 260,000).2
Every check, in the order it ran.
No highlights reel. These are six checks in a row, as they came out of the run. Most retracted papers are small and obscure, and many rows are clean.
| Loading the latest checks… |
Open the full feed → · Data: feed.json
Power
You can verify every number yourself. The census data, method and code behind Campaign One are public, the code under the MIT license.
Love
Built to repair the record, not to shame people. Every result free to everyone, with the census code and data open.
Sound Mind
Flags, not verdicts. No outrage, no villains. Just the count, the evidence and the limits.
- Whole sets only. We never pick targets.
- Method published before the run. Raw results published with it.
- A flag is not an accusation. A citation error is never called fraud.
- Anyone named gets to reply. The reply is attached to the record.
- No funder pays more than about a quarter of a disputed campaign.
- A campaign that misses its finish twice loses its funding. Unless a grant’s terms say otherwise, the unspent money stays with SaneCite’s fiscal sponsor for its other public research.
- Our own numbers get checked too. Every correction is dated and made in public.
Pay for the reading.
The machine has read 55,688 citations. The check by people that confirms or corrects those numbers is designed, registered and ready to run: two readers with biomedical training, 200 citations drawn at random, a pass mark fixed in advance. Gifts will pay the readers, and once SaneCite has a nonprofit fiscal sponsor, every gift and every spend is published.
- Hsiao T-K, Schneider J. Continued use of retracted papers: temporal trends in citations and (lack of) awareness of retractions shown in citation contexts in biomedicine. Quantitative Science Studies 2021. 722 of 13,252 post-retraction citation contexts in PubMed Central acknowledged the retraction.
- SaneCite Campaign One, Phase 1: first run 13 Oct 2026; corrected count 16 Oct 2026 (paperwork retractions left out). Retraction Watch database via Crossref (67,891 retraction records, 21 Oct 2026). Citations from OpenAlex counts_by_year; a citation counts as post-retraction if it falls in a calendar year after the retraction year. Exhibit: OpenAlex W2117847125, 2012–2026.