Retracted papers keep getting cited: by researchers writing later papers, and now by AI research tools too. A retraction does not delete a paper, it adds a notice, and that notice does not automatically follow the paper to every place it gets copied, indexed, or summarized. Independent studies on two well known retracted papers found that a large share of the citations written after the retraction never mentioned it. Tests run in 2025 found ChatGPT and several AI literature-search tools doing the same thing. Below: the named cases, the 2025 AI findings, why this keeps happening, and what you can check yourself right now.
Retraction Watch uses the term zombie paper for a retracted paper that keeps circulating: still cited, still discussed, still built on, as if the retraction had never happened. Their own 2020 example is concrete. A 2018 paper in the Egyptian Journal of Anaesthesia cited three already-retracted papers by Yoshitaka Fujii, a researcher Retraction Watch and Science's own news team both describe as one of the most prolific fraud cases the field has caught. One of the three Fujii papers cited had been retracted in 2013, the other two in 2018, the same year the citing paper itself was published. Nobody caught the overlap in review. It surfaced only in 2019, when a reader wrote a letter to the editor, seven years after Fujii's fraud had already become public knowledge in 2012. That gap, a fraud the field already knew about in 2012 still slipping past editors six years later, is what the term is pointing at.
Vague words like "commonly" hide more than they reveal, so here are two named, measured cases rather than an impression.
The first is Andrew Wakefield's 1998 Lancet paper, which falsely linked the MMR vaccine to autism. Ten of Wakefield's twelve co-authors partially retracted its interpretation in 2004, and The Lancet fully retracted the paper in February 2010. Suelzer and colleagues, writing in JAMA Network Open in 2019, traced 1,153 works that cited the paper. Most, 838 of them (72.7%), cited it negatively, to criticize or debunk it, which is itself a legitimate reason to cite a retracted paper. But narrowing to the works published after the 2010 retraction, 142 of 502 (28.3%) did not mention either retraction at all, despite the paper having already been formally withdrawn. Retraction Watch's own writeup covers the same study.
The second is a 2005 randomized trial published in Chest, claiming a benefit from omega-3 supplementation in patients with chronic obstructive pulmonary disease. It was retracted in 2008 after the underlying data were found to be falsified. Bornemann-Cimenti and colleagues, writing in Scientometrics in 2020, looked at what happened to its citations afterward. Of 112 post-retraction citations they could examine in detail, 107 (96%) never mentioned the retraction, eleven years after it happened.
The same pattern showed up again in 2025, in AI tools built specifically for literature search. MIT Technology Review, reporting on September 23, 2025, tested four tools built for research work, Elicit, Ai2 ScholarQA, Perplexity, and Consensus, by asking each about a set of 21 already-retracted medical-imaging papers. None of the four flagged a single retraction on its own. Elicit referenced 5 of the 21 retracted papers in its answers, Ai2 ScholarQA referenced 17, Perplexity 11, and Consensus 18, all without a word about the retraction. When the outlet retested Consensus in August, after the company said it had added more retraction data sources, its count dropped to 5. That drop matters as much as the original number: it shows the failure is fixable, and that any count here is a snapshot of one tool on one date, not a permanent score.
A separate test, run by Mike Thelwall and published in Learned Publishing in 2025, looked only at ChatGPT. Thelwall gave ChatGPT-4o-mini 217 retracted or otherwise flagged papers, deliberately chosen because they had unusually high public visibility, so the model had the best realistic chance of already knowing something was wrong, and asked it to rate each paper's research quality, 30 separate times per paper, 6,510 ratings in total. Not one of the 6,510 answers mentioned a retraction, a correction, or an ethical concern. Close to three-quarters of the papers still received a high quality rating.
Three reasons, and they compound.
First, retraction status is not attached consistently to a paper wherever it lives. Suelzer and colleagues, in a follow-up study covered by Retraction Watch in July 2021, checked 50 journals and found only 39 consistently marked a retraction on the article's own webpage, and only 31 did so consistently on the posted PDF. Databases disagree with each other too: Web of Science prefixes a retracted title with "RETRACTED:", PubMed does not. If the label itself is inconsistent, anything built on top of it, a citation manager, a search tool, an AI model, inherits the same gap.
Second, both citation practice and AI search tools tend to weight what is easy to find, journal name, keyword match, how often something has already been cited, rather than run a dedicated retraction check. None of the tools tested above check retraction status by default. It has to be added as a separate feature, and as the Consensus retest shows, that addition is possible but not automatic.
Third, and this matters for how to read everything above: citing a retracted paper is not always wrong. Retraction Watch's own guidance is direct: it is fine to cite a retracted paper as long as the retraction is disclosed, for example to explain what went wrong or for historical context. The problem is citing it as if it were still live evidence for a claim, with no mention of the retraction at all. Most of the Wakefield citations above were exactly this kind of legitimate, disclosed, negative citation. The AI-tool failures were not: none of the tools disclosed anything.
A few free or low-cost checks exist right now, for different situations.
If you use a reference manager, Zotero flags retracted items in your library automatically, using Retraction Watch's own data, and warns you before you insert a citation to one in a document. EndNote, from version 20.2, checks your library against the Retraction Watch database and groups retracted items so you can see them together. LibKey Nomad, a free browser extension, flags retractions and expressions of concern directly on publisher pages, Google Scholar, and Wikipedia as you browse. scite.ai's Reference Check is a paid tool that scans a manuscript's entire reference list for the same thing, automatically.
University libraries build guides for exactly this question too, evidence it is a routine one, not a rare one: see HKUST's and SMU's own step by step pages on checking a paper before you cite it.
Paste it and get a plain-words answer, sourced to Crossref, in seconds. No account, no manuscript upload.
Run the check at /checkEvery paper Lucent covers gets a live DOI check and a read of Crossref's retraction-related metadata before publication, and what that check found is printed next to the paper, so you can read it instead of taking our word for it. Every issue stays live in the archive with the screen printed beside each paper, readable without subscribing on the free ones, so this is auditable rather than asserted. Every paper we have published passed that check at the time it went out. That is a fact about our own coverage, not a claim about citation behavior in the wider literature, and nothing in this piece is drawn from our own screening: everything above comes from the independent studies and reporting linked throughout.
The honest limit: a reference-checking tool, including ours, only catches what it is asked to check. Lucent's /check confirms one DOI, at the moment you paste it. It does not fix the underlying problem the studies above describe, that retraction status is not reliably attached to a paper across every database, journal website, and PDF mirror it lives on. No single newsletter, checker, or reference manager fixes that at the source. What each of these tools does is give one more person, at one more moment, a chance to catch what the system does not catch automatically. See also what a retraction actually is and the two-minute check for any study.
Every two days: brand-new papers in plain language, each checked for a live DOI and a Crossref retraction flag before you ever see them. Five free every issue.
What is a paper retraction · How to tell if a study is trustworthy · Check a DOI · Latest issue · Privacy