AI
Fabricated Citations in Academic Papers Surge 12x in Two Years
AI hallucinations are introducing nonexistent references into scholarly literature, creating a verification crisis for journals and researchers.
Key takeaways
- Academic papers with fabricated citations increased more than 12-fold from 2023 to 2025, with one in 277 papers containing fake references by early 2026.
- AI tools prioritize plausibility over accuracy, with GPT-3.5 fabricating 55 percent of citations and GPT-4 fabricating 18 percent in a 2023 study.
- Fabricated citations spread through the research ecosystem as other scholars discover and cite them, assuming they've been vetted.
- Journals are updating policies, with JAMA advising against AI-generated references and submission volumes straining volunteer review processes.
- Librarians recommend conventional reference managers like Zotero over generative AI and emphasize human verification of all citations.
A verification crisis is unfolding in academic publishing as AI-generated hallucinations introduce fabricated citations into scholarly literature at an accelerating rate. Recent analysis of 2.5 million academic papers revealed that the proportion containing at least one fabricated reference increased more than 12-fold between 2023 and 2025. By early 2026, one in every 277 papers contained at least one nonexistent citation.
The problem stems from how large language models construct references. When researchers use AI tools—even just to reformat existing bibliographies to match journal style requirements—the systems prioritize plausibility over accuracy, according to Natalie Ornat Bitting, a librarian at the University of North Carolina at Charlotte. A 2023 study found that 55 percent of citations produced by GPT-3.5 and 18 percent from GPT-4 were completely fabricated, despite appearing legitimate with researcher names, journal titles, and digital object identifiers.
Why it matters
Fabricated citations don't just affect papers where they first appear. Once published, these nonexistent references enter the research ecosystem, where other scholars discover them through literature searches and assume they've been vetted. One fabricated citation from a retracted paper remained discoverable through Google Scholar and appeared in 14 additional sources even after the error was identified, according to Bitting.
How AI introduces citation errors
The pathways to error are multiple. Terri Gotschall, a librarian at Rollins College in Florida, explained that AI tools can transform legitimate sources into nonexistent ones during reformatting, introducing incorrect DOIs, publication details, or page numbers that experts struggle to detect. The tools lack contextual understanding, making small changes with significant consequences.
In one example Gotschall cited from her own research, a chatbot correctly summarized a study about information literacy but incorrectly attributed the findings to AI literacy instruction—changing just two words that fundamentally altered the meaning. These errors become particularly dangerous when researchers rely on abstracts or AI-generated summaries rather than reading full texts.
The verification challenge
The surge in AI-assisted submissions is straining academic review processes. The journal Organization Science saw submission rates increase 42 percent from November 2022 to February 2026, with papers estimated to contain 70 percent or more AI-generated content representing the fastest-growing category.
Bitting, who volunteers as a copy editor, verifies that references exist and DOI links lead to correct publications. However, many journals lack resources for this level of checking, and editors and peer reviewers—often academics volunteering their time—face mounting pressure from increased submission volumes.
Journal policies are adapting. JAMA updated its 2026 guidance to advise against using AI to generate or format references, while Elsevier permits AI assistance but emphasizes authors' responsibility to verify every citation.
Practical responses
Academic libraries are focusing on information literacy education, teaching researchers to evaluate sources and independently verify findings. Gotschall recommended conventional reference-management tools like Zotero, Mendeley, and EndNote over generative AI, and pointed to emerging solutions like LibKey Nomad, a browser extension connecting AI-generated references to institutional library materials.
"There's nothing that beats a good human eye reading through it and checking everything," Gotschall said.
These findings were first reported by Government Technology's education coverage.