A communication to The Lancet states that over a three-year period, 4,046 references out of 2,810 scientific journal articles published were completely fabricated and were likely hallucinated by the AI. (Photo: Getty)
getty
Using artificial intelligence or AI for researchers to write scientific papers is currently far from A-OK. The research team recently lancet Over a three-year period, 4,046 of the 2,810 published scientific journal articles had completely fabricated references. Most of these fabrications are likely AI-induced hallucinations, and one might wonder what else was artificial in these papers.
By 2026, approximately 1 in 277 papers will contain a citation that appears to have been fabricated by AI.
To investigate what might be produced by AI, a research team from Columbia University (Maxim Topaza, Nir Roguinb, Pallavi Guptab, Zhihong Zhanga) and the University of Eastern Finland (Laura-Maria Peltonen) turned to AI. They developed an automatic reference validation system. Because reviewing all 2,471,758 articles with 125,615,773 accompanying references from January 1, 2023 to February 18, 2026 in PubMed Central open access would take a significant amount of time without computational assistance.
Comparing citations listed in papers with actual bibliographic records helped identify and flag discrepancies. The researchers used more AI, Anthropic’s large-scale language model known as Claude 3.5 Haiku, to go through everything that was flagged and separate honest errors from pure fabrications. References not found in databases such as PubMed, Crossref, OpenAlex, and Google Scholar are considered fabricated, not fab.
In the first search year (2023), approximately 2828 articles contained at least one fabricated reference per article. In just two years, by 2025, this had already jumped to 1 in 458. Then, for the first seven weeks of 2026, the ratio will be even higher at 1 in 277 papers. This is more than a 12-fold increase in a relatively short period of time.
60% of citations for a paper published in 2025 were likely fabricated by AI
One paper in particular that merited a big F (fabricated) was published in an open-access oncology journal in 2025 and focused on ureteroileostomy. A whopping 60% (18 out of 30) of the references cited in the paper were fabricated. There were other examples of papers in which a significant proportion of references were fabricated. A total of 246 different papers contained three or more fabricated references.
The research team also identified a specific author who published a large number of papers with fabricated references. For example, in 2025, two authors published 11 different articles in one surgical journal, including 15 fabricated references. The research team commonly referred to so-called “paper mills.” These are research groups that mass produce papers. Review articles had the highest rate of fabrication, which was 57% higher than other types of articles (16.7 per 10,000 vs. 10.6 per 10,000).
Using large-scale language models in AI has a high risk of hallucinations
Without speaking specifically to the author, it is difficult to know why these fabricated citations were published in the paper. The authors probably didn’t do it for fun, like let’s write a scientific paper with Taylor Swift and Justin Bieber as co-authors. My suspicion is that most of the hoaxes result from hallucinations by a particular type of AI (large-scale language models, or LLMs for short, in case you don’t have time to say “language models”).
That’s because the rise in fabricated citations coincides with the rise in the use of generative AI platforms such as ChatGPT, Claude, and Perplexity that use LLM. I discussed such hallucinations in Picking a Funny Bone. psychology today. LLMs can be particularly prone to hallucinations because they simply look at what is associated with what in the large amount of data found on the Internet without actually critically determining what is accurate or not in a professional manner.
AI fabrication will become a growing problem in scientific publishing
The problem of fabricated quotations continues. In fact, the situation is likely to get even worse as more researchers rely on AI to write their papers and junk for-profit scientific journals continue to proliferate. The proliferation of scientific journals that aim to make money but somehow expect scientists to review papers for free has exacerbated quality control problems. Many prominent scientists no longer wish to review papers or serve on the editorial boards of these journals. Because, frankly, these jobs are unrewarding and usually completely unpaid jobs.
One potential way to counter such AI hoaxes is through the use of AI. This means developing AI tools, especially AI-generated materials, that can detect inaccuracies. Yes, scientific publishing can become a kind of battle of the machines. But it’s unclear whether and when scientific publishers will be willing to invest in such AI tools. And how accurate such AI tools will ultimately be.
Nevertheless, scientific publishing is probably on the brink of liquidation. Many publishers view scientific journals as cash cows, resulting in a rapid rise in the number of journals charging researchers thousands of dollars for the “privilege” of having their papers published. Meanwhile, as I’ve said before, funding for scientific research continues to dwindle like cheese. forbes. As a result, researchers have fewer and fewer funds to pay these publication fees. Researchers also have less and less time to continue contributing to scientific journals.
All of this will likely encourage more researchers to cut corners and use less me/myself and more AI in their work. That’s before these AI tools have been fully vetted and checked for accuracy. Something has to give to quote the title of a 2003 romantic comedy.
