Amid rising fears of AI slop in scientific analysis, over 51 papers accepted on the prestigious Convention on Neural Info Processing Techniques (NeurIPS) have been discovered to comprise faux, AI-generated citations.
Greater than 100 hallucinated citations have been discovered throughout 51 papers accepted on the annual convention, which has emerged as one of many greatest analysis occasions within the discipline of synthetic intelligence and machine studying (AI/ML), in line with a report revealed by AI detection startup GPTZero.
The corporate stated it scanned a complete of 4,841 analysis papers accepted at NeurIPS 2025 held in December final yr in San Diego California, United States, for each hallucinated citations in addition to AI-generated textual content. Whereas 51 out of 4,841 analysis papers with faux, AI-generated sources will not be statistically vital, NeurIPS LLM coverage considers any hallucinated citations to be grounds for a paper’s rejection or revocation.
“These NeurIPS papers have already been accepted, introduced dwell, and successfully revealed. Since NeurIPS 2025 had an acceptance fee for foremost observe papers of 24.52%, every of those papers beat out 15,000 different papers regardless of containing a number of hallucinations,” GPTZero stated in a weblog submit revealed on Wednesday, January 21.
Since NeurIPS is a gathering of a number of the main minds of AI analysis, having a analysis paper accepted by the convention is critical. Nevertheless, GPTZero’s findings present that even a number of the world’s main AI consultants wrestle to make sure that the AI instruments they use present correct responses.
NeurIPS will not be the one analysis convention grappling with the problem of AI-generated writing and hallucinations.
Over 50 hallucinated citations in papers beneath evaluation for ICLR 2026 have been detected by GPTZero in December final yr. On-line pre-print repositories have been flooded with low-quality, AI-generated analysis papers. A latest evaluation of arXiv submissions discovered that scientists who seemed to be utilizing LLM-powered instruments posted about 33 per cent extra papers than researchers who didn’t use such instruments, in line with a report by The Atlantic.
Story continues beneath this advert
How did GPTZero detect faux citations?
GPTZero stated it used its in-house developed agentic AI device known as ‘Hallucination Examine’ to scan sources cited within the greater than 4,000 NeurIPS analysis papers and flag any citations in a doc that would not be discovered on-line.
The corporate additionally added that the citations flagged by its device as being an AI-generated faux have been confirmed manually by a human. It has referred to such AI-hallucinated citations as ‘vibe citations’.
“We outline a vibe quotation as a quotation that probably resulted from the usage of generative AI. Our definition excludes apparent spelling errors, lifeless URLs, lacking locators, and different errors which can be plausibly human,” the corporate stated.
GPTZero’s Hallucination Examine device has been made accessible for authors to examine their manuscripts for quotation errors — together with frequent points that may happen with out LLM involvement like lifeless hyperlinks or partial titles. It additional stated that its AI Detector device “permits editors and convention chairs to examine for AI-generated textual content and suspicious citations on the similar time, resulting in quicker and extra correct editorial choices.”
Story continues beneath this advert
What’s NeurIPS?
Based in 1987, NeurIPS is a analysis convention dedicated to learning neural networks and the interaction amongst computation, neurobiology and physics. Since neural networks underpin many of the AI methods at this time, the convention has developed into a serious AI occasion.
The thirty ninth version of NeurIPS held in San Diego, California, final yr noticed a record-breaking 26,000 attendees, twice as many as simply six years in the past, as per a report by CNBC.
Between 2020 and 2025, submissions to NeurIPS elevated greater than 220 per cent from 9,467 to 21,575. Every paper is reportedly peer-reviewed by a number of people who find themselves instructed to flag hallucinations. Over time, the organisers of the convention have needed to recruit better numbers of reviewers in an effort to proceed its acknowledged mission of “rigorous scholarly publishing in machine studying and synthetic intelligence.”
Paradoxically, a analysis paper revealed months forward of NeurIPS 2025 was titled ‘The AI Convention Peer Evaluation Disaster’ and recognized the problem of AI-generated, faux citations as a possible drawback on the convention.