Computer-generated content and the AI/ML retraction record, 2018– 2026: a characterization study
This characterization study of 13,502 AI/ML-related retraction records from 2018 to 2026 reveals that computer-generated content is the largest single inferred cause of retractions, accounting for 34.8% of cases and predominantly affecting papers published between 2021 and 2023.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine the world of scientific publishing as a massive, bustling library. For years, librarians (editors and peer reviewers) have been checking books before they go on the shelves to make sure the stories are true and the authors did the work themselves.
This paper is like a report card on the "Return to Sender" pile of that library between 2018 and 2026. Specifically, it looks at the section dedicated to Artificial Intelligence (AI) and Machine Learning.
Here is the story of what the author, Anton Sokolov, found, explained simply:
1. The New Problem: "Ghost Writers"
Before 2022, if a book in this section was returned, it was usually because the author copied someone else's story (plagiarism), lied about the data, or paid a "paper mill" (a factory that writes fake papers) to do the work.
But starting in 2022, something new happened. Large Language Models (like the AI you might chat with) became very good at writing fluent, scientific-sounding text. Suddenly, the library started getting a flood of books written by these "ghost writers."
The study looked at 13,502 papers in the AI section that were pulled off the shelves (retracted). They asked: Why were these books returned?
2. The Big Discovery
The most surprising finding is that the #1 reason for pulling these AI papers was Computer-Generated Content.
- The Numbers: Out of every 100 retracted AI papers, about 35 were pulled specifically because the text was generated by an AI (like a chatbot).
- The Comparison: This is higher than "Compromised Peer Review" (22%), "Plagiarism" (18%), or "Editorial Mistakes" (13%).
Think of it like a bakery. For years, the main reason bread was thrown out was that it was burnt or stale. But suddenly, the bakery started finding that the dough itself was made by a robot that didn't know how to bake. That became the most common reason for throwing bread away.
3. The "Precision" Filter (The Safety Net)
The author was careful not to just count everything loosely. He created a "safety net" with different levels of certainty:
- High Precision: Papers where the title clearly said "AI" or "Machine Learning." (3,243 papers).
- Broad Net: Papers that were just generally about "Computer Science." (13,502 papers).
Even when looking only at the "High Precision" group (the ones definitely about AI), 41.8% were retracted because of AI-generated text. So, the problem isn't just a fluke of loose labeling; it's a real, dominant issue in this specific field.
4. The Timeline: A Sudden Wave
The paper tracks the years like a weather report:
- 2018–2021: Very few AI-generated papers were caught.
- 2022: The wave started.
- 2023: The wave peaked. This was a "tsunami" year, largely because a major publisher (Hindawi/Wiley) did a massive cleanup, pulling thousands of papers at once.
- 2024–2025: The numbers dropped from the peak but remained 10 times higher than they were before 2022.
Most of these bad papers were written and published between 2021 and 2023. They were caught quickly, usually within a year of being published.
5. Who Was Involved?
The study lists the publishers and journals with the most retracted papers.
- Hindawi (a publisher) had the most, accounting for over a third of all the retracted papers in this list.
- However, the author warns: This isn't necessarily a "shame list." Sometimes, a publisher has a lot of retractions because they are actively cleaning up their shelves and catching bad papers, rather than because they are the "worst" publishers. It's like a store with a "Returns" counter: a high number of returns might mean the store is very strict about quality, not just that they sell bad products.
6. What This Study Is NOT
The author is very clear about what this report doesn't tell us:
- It doesn't tell us how many AI-written papers are still on the shelves and haven't been caught yet.
- It doesn't prove that AI researchers are "bad" people.
- It doesn't say AI research fails more often than other fields (like biology or physics). We just don't have the data to compare them yet.
The Bottom Line
This paper is a "characterization study." It's like taking a snapshot of the "Return to Sender" bin for AI papers.
The main takeaway is simple: The era of AI-generated text has arrived in science, and the first major sign of this is a massive spike in retractions where the text itself was written by a machine.
The study concludes that for science to fix this, we need clear, machine-readable labels on these retractions that say, "This was written by AI." Without those clear labels, we can't measure the problem or fix it properly. The "substrate" (the foundation) for future solutions is simply having honest, clear records of what went wrong.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.