GPTZero finds 100 new hallucinations in NeurIPS 2025 accepted papers
TL;DR Highlight
GPTZero scanned 4841 NeurIPS 2025 papers and found 53 with 100+ fabricated citations (hallucinated references) — a serious academic integrity issue.
Who Should Read
Academic researchers, conference organizers, and anyone evaluating whether AI-generated content in scholarly work is detectable and problematic.
Core Mechanics
- GPTZero's AI detection tool scanned the entire NeurIPS 2025 accepted paper corpus (4841 papers) and flagged 53 papers with 100 or more citations that appear to be hallucinated.
- Hallucinated citations are plausible-sounding but nonexistent references — they often have realistic author names, paper titles, and venues but don't correspond to real publications.
- This is a different problem from AI-generated text detection — it's specifically about fabricated scholarly references, which can cascade through the literature when others cite the citing paper.
- The 53-paper figure likely understates the problem — GPTZero's threshold was 100+ hallucinated citations, so papers with fewer fabricated references weren't flagged.
- NeurIPS 2025 acceptance rate is around 25% — if accepted papers have this issue, rejected papers likely have higher rates.
- The academic community has no established process for systematically checking citations for hallucination at scale, making this a systemic gap.
Evidence
- GPTZero published their methodology and a list of flagged paper IDs, enabling community verification.
- Several researchers independently verified a sample of flagged citations and confirmed the hallucination pattern.
- HN discussion was alarmed: academic citation networks are a foundational trust mechanism, and systematic hallucination corrupts that infrastructure.
- Debate about whether the authors knew (intentional misconduct) or didn't know (accidentally included AI-generated reference lists without checking). Both are problematic for different reasons.
- Conference organizers don't have the capacity to manually verify all citations — this points to a need for automated citation verification at submission time.
How to Apply
- If you're writing academic papers with any AI assistance: run every reference through a citation verifier (Semantic Scholar, CrossRef, Google Scholar) before submission.
- For reviewers: spot-check 5-10 citations in every paper you review — hallucinated references are often in the related work section and may not be obvious.
- Conference organizers: consider adding automated citation verification as part of the submission pipeline — tools like GPTZero and Semantic Scholar can flag suspicious references.
- For research teams: establish a policy that every reference must be independently verified before inclusion, regardless of how the draft was generated.
Code Example
# Example of verifying paper existence using Semantic Scholar API
import requests
def verify_citation(title: str) -> bool:
url = "https://api.semanticscholar.org/graph/v1/paper/search"
resp = requests.get(url, params={"query": title, "limit": 1})
data = resp.json()
return data.get("total", 0) > 0
# Usage
print(verify_citation("Attention Is All You Need")) # True
print(verify_citation("Fake Paper by John Doe 2024")) # FalseTerminology
Related Papers
MemTrace: Tracing and Attributing Errors in Large Language Model Memory Systems
RAG, Mem0 같은 LLM 메모리 시스템이 왜 틀린 답을 내는지 자동으로 찾아주는 디버깅 프레임워크
DeepSWE: A contamination-free benchmark for long-horizon coding agents
기존 SWE-bench의 데이터 오염 및 검증 오류 문제를 해결하기 위해 처음부터 새로 만든 코딩 에이전트 벤치마크로, GPT-5.5가 70%로 1위를 차지하고 모델 간 성능 격차가 훨씬 뚜렷하게 드러난다.
Constraint Decay: The Fragility of LLM Agents in Back End Code Generation
LLM 코딩 에이전트는 구조적 제약(아키텍처 패턴, ORM, DB 설계)이 쌓일수록 성능이 급격히 떨어지는 'constraint decay' 현상을 보인다는 연구 결과로, AI 코딩 도구를 프로덕션에 쓰려는 개발자라면 반드시 알아야 할 한계다.
AMEL: Accumulated Message Effects on LLM Judgments
LLM을 자동 평가자로 쓸 때 이전 대화 기록의 긍정/부정 분위기가 이후 판단을 오염시킨다는 걸 75,898개 API 호출로 증명한 연구.
Language-Switching Triggers Take a Latent Detour Through Language Models
8B LLM에 심어진 백도어 트리거가 중간 레이어에서 언어 탐지기를 완전히 속이는 직교 부분공간(orthogonal subspace)으로 숨어 이동한다는 걸 회로 분석으로 밝혀냈다.
Formal Methods Meet LLMs: Auditing, Monitoring, and Intervention for Compliance of Advanced AI Systems
LLM이 규칙을 잘 지키고 있는지 감시하려면 LLM에게 맡기지 말고 LTL(시간 논리 공식) 기반 모니터를 쓰세요.