Columbia researchers combed 500,000 science papers for citations that point to invented sources. In 2023 they found 6.3 fabricated references for every 10,000 papers, about 1 in every 1,600. By 2025 the rate had risen to 15.4 per 10,000.
The first three months of 2026 pushed it to 23.0 per 10,000, or 1 in every 435. Columbia University ran the audit. A peer-reviewed journal published it1.
Fabricated References Climb to 23.0 per 10,000 Science Papers
Flagged fabricated references per 10,000 papers, Columbia's audit of 500,000 science papers, 2023 to the first quarter of 2026
Each column is one period Columbia's team audited; the rate nearly quadruples from 2023 to the first quarter of 2026, which is a partial-year estimate.
Source: EurekAlert (Columbia University Irving Medical Center) [1]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| 2023 | 6.3 per 10,000 |
| 2025 | 15.4 per 10,000 |
| Q1 2026 (estimate) | 23 per 10,000 |
Read that number the easy way and the verdict writes itself: AI made science untrustworthy. Computer science, AI's own field, should carry the worst of it.
Topaz's own numbers argue otherwise. Physics flagged fabricated references at 25.6 per 10,000 papers, the highest rate of the five fields his team measured. Materials science followed at 18.4, computer science and AI at 11.1, engineering at 10.3, and chemistry the lowest at 4.31.
Physics Flags Fabricated References 2.3 Times as Often as AI's Own Field
Flagged fabricated references per 10,000 papers by field, Columbia's audit, 2023 to the first quarter of 2026
Each bar is one field Columbia's team audited; physics returns the highest rate, more than double computer science and AI's rate, and chemistry returns the lowest.
Source: EurekAlert (Columbia University Irving Medical Center) [1]. Chart by Ryan Elliott Dennis.
The numbers behind this chart
| Item | Value |
|---|---|
| Physics | 25.6 per 10,000 |
| Materials science | 18.4 per 10,000 |
| Computer science and AI | 11.1 per 10,000 |
| Engineering | 10.3 per 10,000 |
| Chemistry | 4.3 per 10,000 |
The field closest to the technology caught fabrication at less than half the rate of the field furthest from it.
A separate gap explains more of the failure. Of the 619 papers Topaz's team flagged, two carry a retraction so far1.
IEEE retracted 1,417 conference papers in a single day on Oct. 232. A separate study had traced them to paper mills selling co-authorship slots3. Retractions move in bursts, when an editor finally looks, while fabrication moves on its own calendar every day.
Ivan Oransky cofounded Retraction Watch, the database that logs more than 68,000 retractions. He names three causes for the surge, oldest first. "What I liken it to is a fire, where publish-or-perish has always been there as kindling, paper mills are the gasoline, and AI is the accelerant," Oransky says2. Publish-or-perish, in his account, predates any chatbot by decades.
Amanda Sulicz, IEEE's research-integrity manager, reads the same numbers as a verdict on the technology itself. "AI slop is completely eroding trust in science and every scientific process that we've worked for centuries to create," she says2. Her sentence names AI as the cause. Oransky's sentence names the tenure system that existed before AI, with AI as one more tool inside it.
Luigi Longobardi, IEEE's director of publishing ethics, says detection keeps losing the race it runs. "It feels like building a detection tool today may not be effective with the AI-created paper of tomorrow," he says2. Jenny Mahoney, managing editor of IEEE Access, caught one fake submission only because its reference list overflowed. "It was just that so many hallucinated references immediately cast doubt on the entire work," she says.
IEEE Access alone receives about 60,000 submissions a year, and volunteers check nearly every one. The average number of reviewer invitations sent out per article has grown 50 percent since 20182. Reviewers carry that rise on top of their own teaching and research.
Journals built peer review for a smaller flow of papers, decades ago. AI tools can expand that flow for the price of one more prompt.
René Aquarius is a biomedical scientist at Radboud University. He has spent years sorting through flagged papers at Bioengineered, a Taylor & Francis journal unrelated to IEEE's conferences. The publisher retracted 119 more of them on Oct. 1 and 2, out of more than 1,000 articles under review4.
"Given the amount of papers under investigation and the big publication peak that happened in 2021 and 2022, my guess is that more retractions will probably follow in the future," Aquarius says4. His count runs years behind the papers it checks.
Topaz's own data triggered the alarm, yet he resists a flat verdict. "A paper with one unverifiable reference and a paper with five or more are not the same thing, and editors should not treat them the same way," he says1. He wants publishers to spend scarce reviewing hours on the worst cases first. Longobardi and Mahoney give those same hours free, after their own research ends.
Someone pays for every citation that goes uninspected. A reader who builds on an invented finding pays first, in a patient's chart or a bridge's load calculation. Paper mills profit twice, once from the author who buys a slot and once from AI tools that write the filler cheaply.
Reviewers carry that risk themselves, reading submissions on a Tuesday night after a day the university already billed to teaching.
The verdict sits past the easy headline. AI writes citations fast and cheap, and some name invented sources.
Unpaid reviewers are the system built to catch that, reading only on their own time. Reviewers were already behind before the first AI-written paragraph reached an inbox. Fund that work, or keep losing the papers that slip through anyway.
Catching the next retraction before anyone else does takes time spent watching closely, the resource science's reviewers are short on. The Signal, Ai Lately's daily email, spends that time on AI's biggest stories each morning. Readers see who built what, and who got caught, before anyone else does. Join it below.
Sources
- Staff, "Fake references are not only a medical problem: Columbia Nursing audit finds them across physics, engineering, and computer science journals," EurekAlert (Columbia University Irving Medical Center), Oct. 26, 2026, https://www.eurekalert.org/news-releases/1149101
- Benjamin Skuse, "I Used AI to Fake a Scientific Paper. Here's What Happened," IEEE Spectrum, Oct. 27, 2026, https://spectrum.ieee.org/ai-generated-scientific-papers
- Staff, "Weekend reads: Mega-journal on hold at PubMed; 1,400 retractions in one day; 'I Used AI to Fake a Scientific Paper'," Retraction Watch, Oct. 31, 2026, https://retractionwatch.com/2026/10/31/weekend-reads-mega-journal-on-hold-at-pubmed-1-billion-lawsuit-against-elsevier-tossed-i-used-ai-to-fake-a-scientific-paper/
- Staff, "Bioengineered retracts 119 more papers as investigation continues," Retraction Watch, Nov. 3, 2026, https://retractionwatch.com/2026/11/03/bioengineered-retracts-119-more-papers-as-investigation-continues
- Staff, "Fraudulent citations, blamed on AI hallucinations, are becoming more common in research papers," STAT News, May 7, 2026, https://www.statnews.com/2026/05/07/lancet-study-finds-steep-rise-fraudulent-citations-academic-papers/
