By ByteBot
GPTZero analyzed over 4,000 research papers from NeurIPS 2025—the world’s most prestigious AI conference—and discovered something alarming this week: 100+ AI-hallucinated citations that slipped through peer review, spanning 53 accepted papers. The ultimate irony? AI researchers, the very people building and studying these systems, were fooled by fake references generated by tools like ChatGPT and Claude. These fabricated citations included nonexistent authors, fake paper titles, dead URLs, and believable-sounding chimeras combining elements from multiple real papers.
With a 24.52% acceptance rate, each of these papers beat out 15,000+ competitors despite containing verifiable fabrications. Moreover, this isn’t isolated: GPTZero found 50+ similar hallucinations at ICLR 2026. If the world’s top AI researchers can’t spot AI hallucinations in their own papers, the problem is systemic. Developers implementing algorithms based on these citations are building on quicksand.
100+ Hallucinations Slip Through Elite Peer Review
GPTZero’s investigation revealed approximately 2% of accepted NeurIPS 2025 papers contained at least one hallucinated citation. Each paper underwent review by 3+ peer reviewers who missed the fabrications entirely. Furthermore, the types of fakes ranged from blatant to sophisticated.
Some were laughably generic: author names like “John Doe and Jane Smith” paired with invented arXiv IDs (2305.XXXX). Others were far more convincing—real author names paired with fake paper titles, or correct titles assigned to wrong journals with fabricated DOIs. The most insidious were subtle modifications: starting from a real paper but expanding author initials into guessed first names, paraphrasing titles, or swapping publication venues just enough to become unverifiable.
The peer review process, long considered science’s quality gate, failed spectacularly. Under academic norms, even a single fabricated citation is grounds for rejection. Citations anchor papers in existing research and demonstrate that authors have actually read the work they cite. These hallucinations break the knowledge chain entirely.
AI Writing Without Verification: A Perfect Storm
The NeurIPS fiasco exposes a perfect storm of AI over-reliance and peer review strain. Researchers used AI assistants to draft sections—particularly introductions and related work—without verifying the generated citations. Meanwhile, peer reviewers, overwhelmed with reviewing 3+ papers each under tight deadlines, assumed authors had verified references and didn’t spot-check citations themselves.
Here’s the kicker: up to 17% of peer reviews at major computer science conferences are now AI-written, meaning AI-generated papers may be reviewed by AI-assisted reviewers—a double-AI failure loop. Studies show that “when it’s close to the deadline, probability of people using AI seems to increase significantly.” Both authors and reviewers cutting corners created the perfect environment for errors to compound.
OpenAI’s own research explains why this happens: “Language models hallucinate because standard training and evaluation procedures reward guessing over acknowledging uncertainty.” LLMs learn citation formatting patterns but don’t have access to actual paper databases. When asked to cite supporting evidence, they generate text that looks like a citation without verifying existence. Consequently, the result is believable fabrications that fool even experts.
Related: Cognitive Debt: MIT Study Shows ChatGPT Harms Your Brain
This isn’t about individual researchers cutting corners—it’s about system design failing to keep pace with AI adoption. The issue is cultural: treating AI output as truth without verification. The irony is stark: AI researchers, who understand hallucinations better than anyone, still fell victim because convenience trumped verification.
Cascading Errors Break the Knowledge Chain
The impact extends far beyond academic embarrassment. Developers implementing algorithms often rely on citations in papers to understand foundational work. When those citations are fake, the entire implementation may be based on nonexistent research. The problem cascades: other researchers citing these 53 papers may propagate the hallucinations into their own work, creating a chain of misinformation.
The NeurIPS board’s response reveals a troubling disconnect: “Even if 1.1% of papers have one or more incorrect references due to LLMs, the content of the papers themselves are not necessarily invalidated.” This dismissive stance ignores the cascading trust problem. In 2025, NeurIPS reviewers were explicitly instructed to flag hallucinations—yet 100+ still got through.
For developers: if you’re implementing an algorithm from a NeurIPS paper and its citations include fake references, you’re building on shaky foundations. For the research ecosystem: trust is currency in science. When 2% of papers at the top conference contain fabricated citations, the entire peer-review model is in question. If AI researchers can’t police AI output in their own domain, what hope is there for other fields adopting AI tools?
Verification Tools and Cultural Shift Needed
The solution isn’t banning AI—it’s enforcing verification standards. However, GPTZero’s Hallucination Check tool, which caught these errors, works by searching open web and academic databases (Google Scholar, PubMed, arXiv) to verify each citation’s existence. The tool claims 99%+ accuracy, with every flagged citation manually verified by human experts to eliminate false positives.
Researchers can use similar tools like CiteSure or Citely, or implement manual verification workflows. The process takes approximately 30 seconds per citation: copy citation into Google Scholar, verify exact title and authors match, check DOI/URL actually resolves, confirm the paper says what you claim. For 50 citations, that’s 25 minutes—worth it to avoid NeurIPS-level embarrassment.
But the bigger fix is cultural: treating AI output as drafts requiring verification, not final truth. Conference-level solutions include mandatory citation checks pre-review, updated reviewer guidelines to spot-check 5 random citations per paper, and post-acceptance scans before final publication. Technical solutions exist; adoption lags behind AI tool proliferation.
No one, not even AI experts, is immune to AI tricks. Verification must become non-negotiable. If you use AI to draft anything with citations, budget 10-15% of your time for verification. Deadline pressure is not an excuse—it’s exactly when you’re most likely to make mistakes.
Key Takeaways
- GPTZero found 100+ AI-hallucinated citations across 53 papers at NeurIPS 2025, spanning approximately 2% of accepted papers despite 3+ peer reviewers per paper
- The problem is systemic: ICLR 2026 has 50+ similar hallucinations, and 17% of peer reviews at major CS conferences are now AI-written, creating a double-AI failure loop
- AI researchers fell victim to the same hallucinations they study because convenience trumped verification—if experts can’t spot AI tricks, no one is safe
- Technical solutions exist (GPTZero, CiteSure, manual workflows), but the real fix is cultural: treat AI output as drafts requiring verification, never final truth
- For developers and researchers: budget 10-15% of writing time for citation verification; for 50 citations, that’s 25 minutes to avoid building on nonexistent foundations
The NeurIPS incident is a warning shot: convenience doesn’t justify breaking the knowledge chain. Verification is non-negotiable, not optional.
—
