What happens to research after it is retracted.
We matched every retracted paper with a DOI to the citations it received after its retraction, then three AI models read every citing passage we could reach. The set defines itself: a paper is in because a journal retracted it, not because we picked it. Every chart here is free to share with credit to SaneCite (sanecite.com).
Three AI models read every citation we could reach.
Every citation made after retraction to a PubMed-indexed paper retracted from 1990 to 2023, wherever the citing paper is open access: 55,688 citations to 12,864 retracted papers. All three models read each one, and blind samples measure how often they are wrong. Every figure here is 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Out of every 100 citations made after a paper was retracted, 41 treat it as support, 56 mention it in passing, and just 3 flag it as retracted or unreliable.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).
Open “Show the numbers” under the chart for the confidence intervals and the estimate corrected for reader error. Search every citation →
What the census found.
Each chart has its own page, a download that carries its source, a caption to paste and the data.
34.9% of citations still treat a retracted paper as support more than ten years later.
In the first two years after a retraction it is 43.6%. Only a few in a hundred ever flag the paper as retracted or unreliable.
Treated it as support, by years since the retraction
Years after retraction · small numbers = citations that year · tallest: 16,108 in year 1
- Treated it as support
- Mentioned it in passing
- Flagged it as retracted or unreliable
Each bar mixes papers retracted in different years. Recent retractions can only appear in the early bars, so part of the drop comes from which papers are old enough to count.
Bars mix papers retracted in different years: part of the drop is which papers are old enough to count.
Source: SaneCite Phase 2 census: 51,626 citations from open-access papers citing retracted biomedical papers (PubMed-indexed, retracted 1990–2023); the year bars show the 51,370 made within 20 years. 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
| Years after | Citations | Treated it as support (95% CI, %) | Flagged it (95% CI, %) |
|---|---|---|---|
| 1–2 | 25,915 | 43.6% (42.6–44.4) | 2.7% (2.1–3.8) |
| 3–5 | 13,700 | 39.1% (38.0–40.2) | 3.4% (2.9–3.9) |
| 6–10 | 8,277 | 37.4% (35.9–39.0) | 3.8% (2.8–5.4) |
| 11+ | 3,734 | 34.9% (32.4–37.2) | 6.3% (3.2–12.0) |
Bars count citations made that many calendar years after the retraction year (citations in the retraction year itself are not counted). The 256 citations made 21 to 36 years after are in the table and the CSV, not drawn. 95% confidence intervals (the likely range) come from re-running the numbers 400 times, drawing whole retracted papers at random (a cluster bootstrap).
2 in 5 citations still treat a retracted paper as support, even when the citing paper's published reference list shows the retraction.
That is 40.4% of 3,628 such citations, about the same as for all citations (40.7%). Only 12.9% flag it as retracted or unreliable in the text.
Treated it as support
Only 12.9% of the marked citations flagged it as retracted or unreliable in the text.
Source: SaneCite Phase 2 census: 51,626 citations from open-access papers citing retracted biomedical papers (PubMed-indexed, retracted 1990–2023). Dot = estimate; line = 95% CI. 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
| Group | Citations | Treated it as support (95% CI, %) | Flagged it (95% CI, %) |
|---|---|---|---|
| Reference list says “retracted” | 3,628 | 40.4% (37.4–43.5) | 12.9% (7.6–18.0) |
| All citations | 51,626 | 40.7% (39.9–41.5) | 3.3% (2.6–4.3) |
“Reference list says retracted” means the word “retract” appears in the citing paper's own reference entry for the work. Most entries carry a clear mark: a title prefix such as “RETRACTED:” (2,664) or a retraction note (523). The other 441 mention it in other formats. On the 3,187 clear marks alone, 41.5% (1,322) cite the work as support. The mark may come from the authors' reference software or from the journal. 95% confidence intervals (the likely range) come from re-running the numbers 400 times, drawing whole retracted papers at random (a cluster bootstrap).
Only 1.9% of citations to papers retracted for misconduct flag them as retracted or unreliable.
For papers retracted for error it is 6.5%. Both are still cited as support about 4 times in 10 (41.2% and 40.0%).
Flagged it as retracted or unreliable, by reason for the retraction
Misconduct and error are still cited as support about as often: 41.2% and 40.0%.
Source: SaneCite Phase 2 census: 51,626 citations from open-access papers citing retracted biomedical papers (PubMed-indexed, retracted 1990–2023); 4,044 more, to paperwork retractions, are reported apart. Dot = estimate; line = 95% CI. 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
| Group | Citations | Treated it as support (95% CI, %) | Flagged it (95% CI, %) |
|---|---|---|---|
| Misconduct, fabrication or duplication | 31,947 | 41.2% (40.0–42.3) | 1.9% (1.2–3.1) |
| Error or unreliable results | 15,715 | 40.0% (38.9–41.1) | 6.5% (5.1–8.1) |
| Other or no reason given | 3,964 | 40.2% (38.0–42.3) | 2.6% (2.0–3.5) |
| Paperwork retractions (no fault found in the science) | 4,044 | 38.0% (33.9–42.9) | 1.4% (1.0–1.9) |
Reason groups follow the Retraction Watch reason codes. Paperwork retractions (out of date, replaced, moved journal, publisher duplicate) find no fault in the science and are reported apart. 95% confidence intervals (the likely range) come from re-running the numbers 400 times, drawing whole retracted papers at random (a cluster bootstrap).
In every subject, 31% to 43% of citations to retracted papers still treat them as support.
From biology and medicine to the physical and social sciences, business and the humanities. Only 1.7% to 4.1% flag the paper as retracted or unreliable.
Treated it as support, by subject of the retracted paper
A paper can have more than one subject; its citations count under each. Humanities rests on 194 citations, so it is the least certain.
The registered grouping: each paper sorted by its whole set of subject codes
Source: SaneCite Phase 2 census: 51,626 citations from open-access papers citing retracted biomedical papers (PubMed-indexed, retracted 1990–2023). Subjects as listed by Retraction Watch; the subject breakdown is a later check, not in the registered plan. Dot = estimate; line = 95% CI. 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
By subject (a later check, not in the registered plan). A paper with several subjects counts under each, so the groups overlap.
| Group | Citations | Treated it as support (95% CI, %) | Flagged it (95% CI, %) |
|---|---|---|---|
| Basic life sciences | 37,031 | 42.5% (41.6–43.4) | 2.9% (2.2–3.9) |
| Health sciences | 30,858 | 39.3% (38.2–40.3) | 4.0% (2.9–5.6) |
| Social sciences | 1,910 | 36.0% (33.1–39.0) | 4.1% (2.5–6.7) |
| Humanities | 194 | 35.1% (22.6–45.8) | 3.6% (0.4–8.2) |
| Environment | 907 | 34.1% (28.3–40.0) | 3.6% (0.8–7.8) |
| Physical sciences | 2,391 | 32.2% (29.2–35.2) | 3.2% (1.9–5.1) |
| Business and technology | 1,858 | 31.4% (28.7–34.2) | 1.7% (1.0–2.6) |
The registered grouping sorts each paper by its whole set of Retraction Watch subject codes: basic life sciences (BLS) only, health sciences (HSC) only, both, or a set that includes any other subject (such as chemistry or psychology, often alongside BLS or HSC).
| Group | Citations | Treated it as support (95% CI, %) | Flagged it (95% CI, %) |
|---|---|---|---|
| Basic life sciences only | 18,384 | 44.4% (42.9–45.6) | 2.1% (1.6–2.8) |
| Both basic life and health sciences | 16,730 | 41.3% (39.7–42.8) | 3.6% (2.2–6.1) |
| Health sciences only | 11,433 | 37.2% (35.6–38.9) | 4.9% (3.4–7.0) |
| Includes another subject | 4,997 | 34.0% (31.8–35.9) | 3.7% (2.6–5.0) |
| Business and technology only (not grouped) | 82 | 20.7% (9.8–31.9) | 0 of 82 |
The registered rule's code pattern skipped the business and technology code (B/T), so 82 citations to papers coded only B/T are left out of its groups; they are in the table, not drawn, and counted in the subject breakdown. 95% confidence intervals (the likely range) come from re-running the numbers 400 times, drawing whole retracted papers at random (a cluster bootstrap).
- Papers already in press do not explain it.Leaving out the 1,483 citing papers known to have been received before the retraction moves the overall rate by 0.1 point (received dates are known for about 43%).
- Reviews cite it as support most often.More than half the citations in reviews, meta-analyses and systematic reviews cited the retracted paper as support. An exploratory check, done by AI, found 472 meta-analyses, systematic reviews and guidelines that counted a retracted study as data.
- Most citations are passing ones.55.9% were one reference in a bundle, background, or a description. They are counted, but not as support.
The chart from the separate citation-accuracy test sits with that test on the Checker page.
Copy a line, and its source comes with it.
- 40.7% of citations to retracted papers still treat them as support.21,037 of 51,626 post-retraction citations (95% CI 39.9–41.5%). 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%). Phase 2 census, data as of 14 Oct 2026, pending human review.
- Only 3.3% of post-retraction citations flag the paper as retracted or unreliable.1,726 of 51,626 citations (95% CI 2.6–4.3%), judged from the text around each citation. Phase 2 census, data as of 14 Oct 2026, pending human review.
- More than ten years after a retraction, 34.9% of citations still treat the paper as support.43.6% in the first two years (25,915 citations); 34.9% after more than ten (3,734). Phase 2 census, data as of 14 Oct 2026, pending human review.
- Citations to papers retracted for misconduct flag the problem 1.9% of the time, against 6.5% for error or unreliable results.Both groups are cited as support about as often: 41.2% and 40.0%. Phase 2 census, data as of 14 Oct 2026, pending human review.
- 2 in 5 citations still treat a retracted paper as support even when the citing paper's published reference list shows the retraction.1,465 of 3,628 citations (40.4%, 95% CI 37.4–43.5%). Phase 2 census, data as of 14 Oct 2026, pending human review.
- About 1 in 14 citations badly misreport the paper they cite.66 of 915 random citations from open-access PubMed Central papers (7.2%, 95% CI 5.7–9.1%); 20.3% have some error. First citation-accuracy test, 22 Oct 2026, pending human review: measured against AI readers, not people.
- Citations that credit a specific finding badly misreport it 9.7% of the time.40 of 413, against 5.6% for background citations. Not in the test's plan; the same order held in an earlier set of 500. First citation-accuracy test, 22 Oct 2026, pending human review: measured against AI readers, not people.
- Nearly two-thirds of retracted papers were cited again after the retraction.Phase 1 count: 32,717 of 51,470 retracted papers with a DOI (retracted 2024 or earlier, found in OpenAlex), 264,039 citations in all, paperwork retractions left out. Last updated 25 Oct 2026, 11:02 UTC.
A citation here means one later paper citing one retracted paper, in a year after the retraction. The figures are not corrected for reader error (the corrected estimate →).
Count, read, tell.
A retraction is a journal saying, in public, that a paper should not be relied on. If the record can’t carry that message, it can’t carry subtler ones. And retractions are written down, dated and machine-readable, so the check can be run whole and run again by anyone.
The stakes are real. Reviews and treatment guidance are built from the literature one citation at a time. In one large study of biomedicine, only 5.4% of citation contexts citing a paper after its retraction said it had been retracted.1 Small studies have read samples of the rest; nobody had read them all to see how often a retracted paper is still “cited as support”. That is the job.
- Done
1Count
Match every retracted paper to every citation it got after the retraction. Mechanical, no judgement calls.
Out: 63.6% of retracted papers, nearly two-thirds, were cited again (32,717 of 51,470); 264,039 citations in all, leaving out paperwork retractions (updated 16 Oct 2026 from 286,050). Every check, in the order it ran →
- Census read · pending human review
2Read
For every citing paper with open full text, find 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. Results →
- Lists out · replies printed
3Tell
Write to the authors and journals whose papers most often cite retracted work as support, and to the authors of the retracted papers cited most, with the evidence. Then publish each reply and correction beside its entry, and note where no reply came.
Out: the Lists, with every reply beside its entry (all replies).
How the count works.
- Retractions come from the Retraction Watch database, distributed by Crossref, which records the retraction date (67,916 retraction records). Every source is open, so the data cost nothing.
- Citations come from the OpenAlex graph, by year. A citation counts as post-retraction if it falls in a calendar year after the retraction year, so citations in the same year are left out and the count is low, not high.
- Coverage: 53,368 of the 53,732 retracted works with DOIs, retracted in 2024 or earlier, matched in OpenAlex (25 Oct 2026).
- Paperwork retractions are left out: those that find no fault in the science (withdrawn as out of date, temporary removal, publisher error, replaced at the same DOI). That is 1,855 papers and 23,532 citations (25 Oct 2026), by the same rule the Phase 2 census used. Updated 16 Oct 2026: the first count (286,050, run of 13 Oct) included them.
- Accuracy of the count: 200 counted citations, drawn at random and checked against the citing papers’ own reference lists (16 Oct 2026), were about 96% real; the rest were wrong links, a retraction notice, a whole journal issue, or duplicate copies of one paper. The true total is about 254,000 (likely 247,000 to 260,000).
- A post-retraction citation is a flag, not an accusation. Authors often never saw the notice. The audit measures that failure of the system, not of scientists.
- Anyone named in the results can reply, and the reply is attached to the record.
The data behind every number on this site: results.json and feed.json.
How the reading was run.
| Frame | 22,553 retracted papers in Retraction Watch (indexed in PubMed, retracted 1990–2023). 16,019 (71%) were cited in a later year: 130,315 citations. (Phase 1 counts a larger set through OpenAlex, so its share differs.) |
|---|---|
| Citations | 58,358 citations had open full text in Europe PMC. The citing passage was found in 96%; notices were set aside, leaving 55,688 citations read, 43% of all later citations. |
| Reading | Three Cloudflare models (Jev, Clef, Gemma) answer two questions about each passage, and a fixed rule turns the answers into a label. The models are never told the paper was retracted. |
| Check | On a blind random sample of 150 citing papers, labelled by a separate AI reader, the final version of the panel agreed on 94.7%. An earlier version agreed on 92.3% of a first sample of 300; the errors that sample showed were fixed before the census figures were made. On 8,944 citations also checked by Hsiao and Schneider, whose labels were screened by keyword and confirmed by hand, it agreed on 97.7% about whether the retraction was acknowledged. |
| Robustness | Six alternative cuts of the data, from first passage only to citing papers received after the retraction, all give 36–41%. |
| Reading more | After the census, 7,350 more citations were read (6,725 of them judged), mostly from papers that are not open access: free legal copies, and author manuscripts from PMC’s text-mining dataset. Across all 58,339 judged, the rate is 40.1% (95% CI 39.3–40.9%), against 40.7% for the open-access census, and 48% of all later citations have now been read. The two larger new sets on their own run lower (33–39%) and come from a different mix of papers. Registered as protocol amendments 6–11 in the census repository; the figures on this site keep the open-access census. |
| Verify | Cheap to verify: every call is published with its passage, and the public code re-runs it. |
| Data | Search every citation · citations.csv · census.json · codebook.md · protocol and code |
Updated 20 Oct 2026: 12 census citations were matched to the wrong Retraction Watch record. Without them the census read 55,676 citations and judged 51,614. The headline rates (40.7% cited as support, 3.3% flagged) are unchanged; a few subgroup figures move by about 0.1 point (first two years 43.6% → 43.5%; flagged after ten years 6.3% → 6.4%), and the small year-20 bar by 1 point. The pages and the downloadable files keep the published counts, with the 12.
Limits: open-access biomedical papers only, so 57% of later citations are counted but not read; reweighting for them moves the rate by about one point. Labels judge the citing text, not whether the citing paper's conclusions depend on it, and “cited as support” also covers using a method or saying results agree. The pilot that came first (1,002 sampled citations, 38%) keeps its data: results.json · labels.csv.
The human check is ready to run.
So far, AI models have checked the Phase 2 reading. The check by people is designed, built and registered, and gifts will pay its readers.
- What it is.Two people with biomedical training, working alone, each read the same 200 citations (254 passages), drawn at random from all census citations, including the paperwork retractions reported apart. They are not shown the machine’s answer, and are not told whether the cited paper was retracted.
- The rule is fixed in advance.If the machine agrees with either reader on fewer than 85% of the 200, the headline is corrected using the human labels. The 85% rule was set in the original protocol; how it is scored, and that either reader can trigger it, was registered on 14 October 2026 (amendment 4), before anyone was recruited.
- What is ready.The random draw, the reading sheets with their instructions, and the scoring code are built and tested. Each reader needs about 3 to 4 hours.
- What it changes.With it, the 40.7% is confirmed or corrected in public. The result is published either way. Back the human check →
Everything is open.
- How to citeJoseph S. SaneCite retraction census: post-retraction citations to 22,553 retracted biomedical papers (protocol, code and data), version 1.2.0. Zenodo; 2026. doi:10.5281/zenodo.23539192. The latest version is always at doi:10.5281/zenodo.23358099. The 22,553 in the title is every retracted paper the census checked, not only those with a readable later citation.
- Check any numberEach chart links to its own CSV, and the census code and data are public, the code under the MIT license, on GitHub. The accuracy chart’s data is in the test result, and every citation can be searched, with the sentence that was read.
- Look up one paperSearch a title, DOI or PubMed ID on the checker page and open it to see how later papers used it, year by year.
- QuestionsWrite to letters@sanecite.com. Found a wrong call? Send the citing paper's PMC number and it goes into the next release.
- 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;2(4):1144–1169. 722 of 13,252 post-retraction citation contexts in PubMed Central acknowledged the retraction.