Yet Canadian startup GPTZero analyzed more than 4,000 research papers accepted and presented at NeurIPS (Neural Information Processing Systems) 2025 and says it uncovered hundreds of AI-hallucinated citations that slipped past the three or more reviewers assigned to each submission, spanning at least 53 papers in total. The hallucinations had not previously been reported.

From fully made-up citations to subtle changes

In some cases, an AI model blended or paraphrased elements from multiple real papers, including believable-sounding titles and author lists, the company says. Others appeared to be fully made up: a nonexistent author, a fabricated paper title, a fake journal or conference, or a URL that leads nowhere.

In other cases, the model started from a real paper but made subtle changes—expanding an author’s initials into a guessed first name, dropping or adding coauthors, or paraphrasing the title. Some, however, are plainly wrong—citing “John Smith” and “Jane Doe” as authors, for example.

When reached for comment, the NeurIPS board shared the following statement: “The usage of LLMs in papers at AI conferences is rapidly evolving, and NeurIPS is actively monitoring developments. In previous years, we piloted policies regarding the use of LLMs, and in 2025, reviewers were instructed to flag hallucinations. Regarding the findings of this specific work, we emphasize that significantly more effort is required to determine the implications. Even if 1.1% of the papers have one or more incorrect references due to the use of LLMs, the content of the papers themselves are not necessarily invalidated. For example, authors may have given an LLM a partial description of a citation and asked the LLM to produce bibtex (a formatted reference). As always, NeurIPS is committed to evolving the review and authorship process to best ensure scientific rigor and to identify ways that LLMs can be used to enhance author and reviewer capabilities.”

Edward Tian, cofounder and CEO of GPTZero, which was founded in January 2023 and raised a $10 million Series A round in 2024, told Fortune the NeurIPS analysis came just weeks after the company uncovered 50 hallucinated citations in papers under review for another top AI research conference, ICLR, which will be held in Rio de Janeiro in April. In that case, the papers had not yet been accepted—but the bogus citations had already slipped past peer reviewers. Tian said the ICLR conference has hired the company to check future submissions for fabricated citations during peer review.

Errors appeared in papers accepted and presented at NeurIPS

According to Tian, the NeurIPS findings are even more troubling because the errors appear in papers that were accepted by the conference. In the academic world of AI, “publish or perish” is more than a cliché: Hiring and tenure often hinge on accumulating peer-reviewed publications. Yet under long-standing academic norms, even a single fabricated citation would, in principle, be grounds for rejection. References are meant to anchor a paper in the existing body of research—and to demonstrate that its authors have actually read and engaged with the work they cite.

“It’s definitely a bigger escalation in the sense that these were the first documented cases of hallucinated citations entering the official record of the top machine learning conference,” Tian said, pointing out that since NeurIPS 2025 had an acceptance rate for main track papers of 24.52%, each of these papers beat out 15,000 other papers despite containing one or more hallucinations. “These survived peer review, and were published in the final conference proceeding,” he said. “So it’s definitely a big moment.” 

Around half of the papers with hallucinated citations were papers that were likely to be AI-generated themselves or had a high amount of AI use, he added. “But what we were really focused on in this investigation is the citations themselves,” he said. AI detection tools have often been criticized for false positives in attempting to identify machine-written text. But Tian argued that hallucination detection is a different class of problem, with GPTZero’s tool checking verifiable facts—searching the open web and academic databases to confirm whether a cited paper actually exists. The company says the tool is more than 99% accurate, and for the NeurIPS analysis, every flagged citation was also reviewed by a human expert on GPTZero’s machine-learning team.

Alex Cui, Tian’s cofounder and chief technology officer, said that GPTZero’s hallucination checker tool ingests a paper and then searches across the open web and academic databases to verify each citation—its authors, title, publication venue, and link. If a reference can’t be found, or if it only partially matches a real paper, the system flags it. That’s how it catches cases where an AI model starts from a real paper but adds authors who don’t exist, alters the title, or invents a publication. 

“Sometimes, even when there is a match, you’ll find that they added like five authors who don’t exist to a real paper, so these are mistakes that no human would reasonably make,” he explained. For the NeurIPS investigation, after the automated scan, a member of GPTZero’s machine-learning team manually verified every flagged citation, ensuring the findings aren’t themselves false positives.

The sheer volume of papers makes deep scrutiny difficult

A big part of the challenge is sheer scale. In 2025, the main NeurIPS research track received 21,575 valid submissions—up from 15,671 in 2024 and 12,343 in 2023. Even with thousands of volunteer reviewers, that volume makes deep scrutiny of every paper and its references increasingly difficult.

But while AI has a part in that by making it dramatically easier to churn out conference submissions, Tian said, a flawed paper still carries real reputational risk—for the authors, for the conference that accepted it, and for the companies that hire researchers based on those credentials.  

That’s particularly true for citations, he said, because in modern AI research, citations are a core part of how the field tries to solve issues of reproducibility. “AI results are notoriously hard to reproduce, so citations are important,” he said, to “draw the line between whether that result was reproducible or not,” by letting other researchers trace a result back to something concrete and testable. Hallucinated citations, on the other hand, send readers to something that doesn’t exist.