One in 277 peer-reviewed papers published in early 2026 cites a source that does not exist. Not a typo. Not a formatting slip. A fully invented study, complete with a title, an author list, and a publication date, generated by an AI chatbot and never caught before the paper went to print. If you use AI to help draft reports, proposals, or research summaries at work, that number should worry you more than any headline about job losses.
The clearest evidence comes from two directions at once. A Dutch investigation by De Groene Amsterdammer and the Utrecht University Data School combed through more than 100,000 scientific publications and found 748 references to research that was never conducted, spread across 208 peer-reviewed articles. Separately, a Columbia University Data Science Institute audit of nearly 2.5 million PubMed-indexed papers, led by Maxim Topaz and published as a letter in The Lancet, found the same pattern at global scale.
How fast is the fabrication rate growing?
Fast enough that a three-year-old baseline is already obsolete. The Columbia audit found the rate rising from 1 in 2,828 papers in 2023, to 1 in 458 in 2025, to 1 in 277 in the first seven weeks of 2026 alone, more than a tenfold jump. The Dutch investigation tracked a similar curve closer to home: from 1 in 1,400 in 2023 to 1 in 200 in early 2026, a sevenfold increase. Two separate research teams, two different datasets, the same accelerating trend.
For context, Retraction Watch's own database, which tracks retracted and corrected papers worldwide, now lists more than 64,000 entries, updated daily. AI-fabricated citations are quickly becoming one of its fastest-growing categories.
Here's the thing: these fake citations are not sloppy. They carry the right formatting, they name real researchers, they cite journals that genuinely exist, and they come with plausible publication years. The only thing missing is the paper itself.
The pattern shows up well beyond biomedical research too. At NeurIPS 2025, one of the most competitive AI conferences in the world, 53 accepted papers collectively contained more than a hundred fabricated citations, despite passing peer review that is supposed to be among the strictest in the field.
Why should you care even if you're not a scientist?
Because the same failure mode shows up whenever you use AI to research anything for your job. Picture this: you're pulling together a market analysis, a client proposal, or a policy brief, and you lean on AI to help gather sources. There's a real chance a slice of your bibliography now points to research that was never run, never reviewed, and supports nothing. A client who checks a single citation and finds the study doesn't exist will start doubting the rest of your report too. In regulated industries like healthcare, finance, or legal services, the fallout runs deeper: citations there back up decisions that affect real people.
The legal profession is already living this. According to the Charlotin AI Hallucination Cases Database, more than 1,300 cases have now been documented worldwide where AI-generated hallucinations turned up in court filings, 915 of them in the United States alone. In March 2026, the legal outlet Law360 logged seventeen separate judicial rulings in a single day that flagged AI-fabricated citations in submitted documents. Lawyers who didn't check their AI-generated references got fined and sanctioned.
Why these fake sources slip past everyone
Because AI models are unusually convincing fabricators. Think of it like handing a bibliography to a new hire and asking them to tidy it up. If that new hire quietly slipped in a few invented sources that look professionally formatted, you would only catch it by checking every single citation individually, and that's exactly the step most people skip.
Researchers confronted with fake citations in their own papers described the same pattern: they had used AI for tasks that felt harmless, alphabetizing a reference list, adjusting formatting, translating citations from one language to another. Somewhere in that process, the chatbot quietly added references that were never there to begin with.
In one case, the Dutch investigation found a paper on brain cancer treatment where a proposed therapy was backed by a mix of real and entirely invented sources. Someone reading that paper, a doctor or a fellow researcher, could end up following a treatment direction partly built on research that never happened.
“Everyone was genuinely shocked by what they heard and tried to correct it with the journals immediately.”
Yannick van der Heijden, investigative journalist, De Groene Amsterdammer
Mohammad Hosseini, a research integrity scholar at Northwestern University, points out that sloppy citation practices aren't new. Some researchers were already careless about verifying references before generative AI existed. What's changed is the scale. A problem that used to be occasional is turning structural, because AI lowers the barrier to fabricating a source without the person using it ever realizing it happened.
How a single fake citation snowballs
One of the more unsettling findings is what happens after a fabricated reference gets published once. Investigators found an invented citation that took on a life of its own: other researchers cited it in their own work, and the fake study gained credibility simply by being referenced repeatedly. That's how false scientific consensus forms on a foundation that was never there.
Imagine a report on AI in customer service citing a "study" claiming chatbots resolve 80 percent of queries correctly, a study that never existed. Three more publications pick up that number and cite it onward. Before anyone checks the original source, the "fact" has become common knowledge. The same dynamic playing out in academic publishing can happen in any field where AI-generated text circulates unchecked.
How do you check if an AI-generated citation is real?
Verifying a citation takes about five minutes and can save your report, proposal, or article from resting on fabricated ground. Three steps get you there:
- Search the exact title. Copy the full title of the cited paper and search it in quotation marks on Google Scholar or Semantic Scholar. Zero results usually means the paper doesn't exist. Worth noting: if the title only shows up in other AI-generated text and never on the journal's own site, treat that as a red flag too.
- Check the DOI. If the citation includes a DOI (a unique code starting with 10.), look it up through doi.org. A dead DOI link is an immediate warning sign. No DOI at all for a recent paper from a well-known journal is suspicious in its own right.
- Verify the author and the journal separately. Does the named author exist? Do they actually publish on this topic? Does the journal exist? AI sometimes pairs a real author with an invented journal, or a real journal with an invented author. A quick check of the author's ORCID profile or the journal's website settles it.
If you regularly use AI output for reports or research, build a fixed verification step into your workflow. Not as an afterthought, as part of the process. Make source-checking as routine as a spell check.
What publishers and universities are doing about it
Major scientific publishers admit their existing safeguards no longer hold. Springer Nature and Elsevier have both expanded their research-integrity teams. Elsevier is building AI-detection systems designed to flag fabricated citations automatically before an article gets published.
Bert Seghers, chair of the European Network for Research Integrity Offices (Enrio), is working on new EU-wide reporting standards for AI use in research. "References are only part of the question," he argues. Transparency about how and where AI was used throughout the research process is, in his view, the next step regulators and institutions need to take.
But until those standards exist, the responsibility sits with whoever is using the tool. That applies to scientists, and it applies just as much to you if you use AI for a proposal, a policy brief, or a market analysis. Who's actually checking what your AI produces? That question matters more now than it ever has.
What does this mean for how you use AI at work?
McKinsey's global survey on AI adoption found that 71 percent of organizations now use generative AI regularly in at least one business function, up from 37 percent just three years earlier. That's a lot of teams generating reports, summaries, and citations with tools that are demonstrably prone to inventing sources.
The lowest hallucination rate independent testing has measured across current models sits at 1.8 percent, according to Vectara's ongoing benchmark. For some models, it climbs as high as 95 percent. No model on the market is immune to this failure mode.
The practical takeaway is simple. If you use AI to write text, gather sources, or draft reports, verify every citation it generates before the document goes out the door. Not a spot check, all of them. Five minutes per source is a small price against the damage a report full of unverified AI output can do to your credibility.
The ghost citations turning up in scientific papers carry a warning for anyone using AI for writing or research. Treat every AI-generated citation with the same scrutiny you'd apply to any other source: verify it before you publish.
