Retractions were rare. A paper would be pulled from time to time, after a long and slow investigation, and the news would be limited to the field. That is now an era of the past. In 2023, more than 10,000 papers were retracted by journals in a year, setting a record. In 2024, there was no reprieve, but there was a change in the authors of the retraction. According to numbers shared with Retraction Watch, Springer Nature retracted 2,923 papers in 2024. That was 0.61% of the roughly 482,000 articles it published across more than 3,000 journals, and 61% of the retractions were cleanups of pre-2023 papers. These are not the statistics of a few bad scientists. These are statistics of industrialized fraud and AI text generators, and a system struggling to keep up. Anyone who is in a school system should recognize the tactics of the retraction boom. The same ones we see in student work, but scaled up and turned for profit. What research publishing is learning the hard way should be read closely, because that’s where the habits of these tactics are rooted or corrected.
Two engines: paper mills and paraphrasing
The retraction crisis has two sources, and they feed each other.
First, paper mills. These companies publish research papers, and sell authorship slots to those who need published papers to advance their careers, earn their degrees, or meet their job requirements. A paper mill can fabricate data, create plausible figures, and publish the result in a real journal, with the assistance of an editor who has been compromised. Chemistry World reports that one in fifty published papers now has a paper-mill signature. In 2023, the publisher Wiley shut down 19 journals from its Hindawi portfolio after about 8,000 papers were traced back to paper mills. The magnitude of this problem is hard to overstate: this is not a few bad actors. This is an industry.
Second, text manipulation. This is where the retraction crisis comes into contact with something familiar. Mills and some individual authors use stolen or template text and run it through paraphrasing software to help evade plagiarism checks. The software substitutes common terms for synonyms. The result is a class of errors known as "tortured phrases." Instead of "artificial intelligence," you get "counterfeit consciousness." Instead of "artificial neural network," you get "fake neural organization." Instead of "signal to noise ratio," "flag to commotion proportion." The phrases are meaningless to a specialist, but they fool a naive plagiarism check because the words no longer match the source.
This mechanism sounds an awful lot like the AI "humanizer" and plagiarism paraphrasing tools marketed to students. It is the same mechanism, just with a different price tag and a higher stakes.
Detection is scaling, but it is not enough
The good news in this tale is that detection has come of age. The most widely known implementation is the Problematic Paper Screener, developed by computer scientist Guillaume Cabanac. As reported by The Conversation and TechXplore in early 2025, the screener runs every week through about 130 million publications, running nine detectors to identify tortured phrases, references to retracted work, computer-generated content, and other indicators of fraud. Over 3,500 retractions have been enabled by the screener, and publishers are using it as a standard screen rather than a fallback. That is a major accomplishment, and it confirms an assertion made by Checkmark many times before: detection is useful, and it is not a judgment. The screener does not identify a paper as fraudulent. It identifies a paper that can be examined by a human. Tortured phrases may be the hallmark of a paper mill, but they may also be an indication of a non-native English speaker using a translation tool in good faith. Publishers still need to examine intent, context, and the data before acting. In 2026, Wiley said that if a tortured phrase is found after publication, the paper will be retracted unless an author can demonstrate that the substitutions were made as a result of translation or editing. Note the structure of that rule: the detection raises the issue, and human evidence resolves it. Detection does not solve the more difficult half of the problem, which is that fraudulent results corrupt everything built on top of them. Once a fraudulent paper is cited in a systematic review, or used in a meta-analysis, the contamination persists even after the paper is retracted. Remediation is slower and more costly than the fraud that caused the contamination, which is why so much of the 2024 retraction activity was retroactive.
The disclosure turn
The other response has been to establish rules about how AI can be used. When ChatGPT appeared, scholarly publishing’s first instinct was to answer one question: can a chatbot be an author? The answer, given rapidly and almost universally, was no. The Committee on Publication Ethics, the World Association of Medical Editors, and the JAMA Network all ruled AI tools cannot be named as authors because authorship requires accountability, and software cannot be held accountable. An early editorial, "ChatGPT is fun, but not an author," established the tone, and most journals now follow suit. The next step has been to move from prohibition to disclosure. The International Committee of Medical Journal Editors requires authors to acknowledge AI writing assistance in the acknowledgments and to disclose AI use for data collection or analysis in the methods. Elsevier, in a policy revised in September 2025, requires AI use to be declared in a statement preceding the references, and it prohibits using AI to produce or modify images. The general principle is that AI assistance is not cheating by default. Undisclosed AI assistance is. Transparency is the admission fee. This is a more mature position than a simple ban, and it is the position that schools should be studying. A "no AI" rule is simple to write and very difficult to enforce, and it drives use of AI underground. A disclosure rule raises the more difficult but honest question: did you tell us what you did, and can you back it up?
Why this belongs in the K-12 and higher-ed conversation
Retracting a paper should not be someone else’s problem. It should not be a problem for journal editors or university research offices. That would be a mistake, for three reasons.
First, the pipeline is continuous. The graduate student who bought a paper today learned to relate to sources somewhere earlier. The undergraduate who runs an essay through a paraphrasing tool to beat a similarity score is rehearsing, in miniature, the move that retracts a paper. The norms that govern the top of the pipeline are set at the bottom. What a fifth-grade teacher says about citing a source and what a doctoral advisor says about data provenance are points on the same line.
Second, the tools are identical. Students and paper mills reach for the same category of software for the same reason, which is to make copied or machine-generated text register as original. Checkmark has written about the humanizer arms race, and the research literature is now the highest-stakes example of where that arms race leads. Educators who understand tortured phrases will recognize the tells in student work, and will understand why a single similarity number, high or low, was never the whole story.
Third, the field’s response is a usable template. The research world converged on a workable posture: screen at scale, investigate with humans, require disclosure, and hold people rather than tools accountable. That maps cleanly onto a classroom. A detection score starts a conversation instead of ending one. An honest AI-use policy names what is allowed and asks students to say what they did. Assessment design that values process, drafts, revision history, and the ability to defend an argument out loud is harder to fake than a finished product, whether the faker is a sophomore or a paper mill.
The takeaway
The retraction wave is the phenomenon that occurs when a system, in the interest of a measurable metric (published papers), allows shortcuts to outpace monitoring. The correction that is currently taking place is not the product of a single clever detector. It is a mix of improved screening, more transparent disclosure rules, and renewed demands that some named human be held accountable for the work.
Schools are doing the same thing on a small scale and they are doing it today, with the students who will be writing tomorrow's research. The lesson from research publishing is not that detection failed, or that AI was responsible for the crisis. It is that integrity thrives when detection, transparency, and human judgment work together, and that it dies when any one of them is expected to carry the load alone. That is the lesson worth teaching before anyone ever submits to a journal.
Sources
- Springer Nature retracted 2,923 papers last year, Retraction Watch (Feb 17, 2025)
- Problematic Paper Screener: Trawling for fraud in the scientific literature, The Conversation (Jan 2025)
- Problematic paper screener: Trawling for fraud in the scientific literature, TechXplore (Jan 2025)
- AI tools combat paper mill fraud in scientific publishing as peer review system struggles, Chemistry World
- Publisher AI Policies and Disclosure Rules: A Guide for Authors, Enago
- Prevalence and Trends in Global Retractions Explored Through a Topic Lens, arXiv (2025)
- Detection of tortured phrases in scientific literature, arXiv