The Growing Threat of AI-Generated Content in Preprint Repositories
Preprint servers have long been a cornerstone of modern scientific communication, allowing researchers to share findings rapidly without the delays of traditional peer review. Platforms like arXiv, bioRxiv, and medRxiv enable academics to disseminate preliminary results, solicit feedback, and establish priority in discoveries. However, a surge in artificial intelligence (AI)-generated manuscripts—often low-quality, nonsensical, or outright fabricated—has forced these repositories to overhaul their moderation practices. This shift is particularly acute in the United States, where universities rely heavily on preprints for cutting-edge research in fields like computer science, physics, and biomedicine.
The problem escalated dramatically following the widespread availability of large language models (LLMs) such as ChatGPT and its successors. What began as a tool for writing assistance has morphed into a source of 'junk science,' flooding servers with papers lacking originality or scientific merit. Moderators now grapple with distinguishing genuine work from AI slop, balancing openness with the need to maintain credibility.
A Brief History of Preprint Servers and Their Role in US Academia
Preprints emerged in the 1990s with arXiv, founded at Cornell University in 1991 to serve physicists. The model spread to biology with bioRxiv in 2013 and clinical research with medRxiv in 2019, both operated by Cold Spring Harbor Laboratory. During the COVID-19 pandemic, preprints proved invaluable for sharing vital data quickly, accelerating vaccine development and public health responses.
In the US, where research output leads the world, preprints are integral to university workflows. Faculty at institutions like MIT, Stanford, and Harvard use them to claim priority, collaborate, and build citation records. Yet, this openness has vulnerabilities, exacerbated by AI's ability to mimic scientific prose convincingly.
The Surge: Statistics on AI Spam in Preprints
arXiv receives over 25,000 submissions monthly, with computer science seeing hundreds of low-quality reviews and position papers alone—many suspected AI-generated. Moderators report a 'flood' since late 2025, prompting policy changes. bioRxiv and medRxiv note similar rises, with titles like 'Self-Experimental Report: Emergence of Generative AI Interfaces in Dream States' raising red flags.
Studies indicate AI boosts paper production, particularly for non-native English speakers, but quality suffers. One analysis found AI-assisted papers proliferating on Google Scholar, mimicking legitimate research. In the US, this junk dilutes the signal for genuine breakthroughs from university labs, complicating literature reviews and funding decisions.
arXiv's Bold Policy Shift: Banning Certain Paper Types
In response, arXiv—US-based at Cornell—enforced stricter rules in January 2026. Computer science reviews and position papers now require proof of prior peer review elsewhere; otherwise, rejection. This targets AI slop resembling annotated bibliographies without analysis. Endorsement systems, needing institutional emails and prior arXiv history, were tightened to filter spam.
Moderator Kat Bohner highlighted LLMs evading plagiarism checks, producing convincing but valueless content. US computer science departments at places like UC Berkeley and Carnegie Mellon, heavy arXiv users, applaud the move but worry about overreach blocking innovative work.
bioRxiv and medRxiv: Hands-Off Yet Vigilant Approaches
bioRxiv and medRxiv maintain lighter moderation, focusing on provenance over quality. They reject if fabrication is evident but let community feedback handle merit. Pandemic lessons, like rejecting unverified COVID treatments, inform current AI checks. John Inglis, co-founder, emphasizes post-publication scrutiny.
For US biomedical researchers at Johns Hopkins or NIH-funded labs, this preserves speed while curbing harm. However, rising paper mill activity—AI-amplified—prompts internal debates on deeper screening.
Detection Tools and Moderation Techniques Evolving Rapidly
Servers employ diverse strategies: ORCID verification, author reputation, formatting checks, and AI detectors. LatArXiv scrutinizes profiles; RINarxiv contacts authors directly. Emerging tools analyze perplexity and burstiness to flag LLM output.
- Endorsement systems (arXiv): Prior authorship required.
- Advisory boards for edge cases.
- Gut checks on titles and abstracts.
- Community reporting for post-upload removal.
US universities are developing in-house detectors, with Stanford piloting LLM classifiers for internal reviews.
Impacts on US Higher Education and Research Ecosystem
US universities face citation pollution, where junk dilutes discoverability. Early-career researchers at state schools struggle amid noise. Funding bodies like NSF scrutinize preprints more, delaying grants. A Nature survey shows 14 moderators noting legitimate rejections, frustrating open science ideals.
Alice Fleerackers (University of Amsterdam, but relevant) faced repeated arXiv rejections despite track record, echoing US complaints. Cornell's arXiv independence push signals institutional strain.
Expert Perspectives: Balancing Openness and Integrity
Natascha Chtena (Simon Fraser): Moderators walk a 'difficult balance.' Fleerackers warns of inclusivity loss. Inglis trusts community vetting. US voices like MIT's call for transparency in AI use disclosure.
Panel discussions at AAAS meetings highlight US labs' push for standardized policies across servers.
Stakeholder Views: From Moderators to Publishers
Publishers like Nature urge disclosure of AI use. Universities train faculty on ethical AI. Paper mills exploit gaps, but AI spam is new scale. Solutions include watermarks on AI text and blockchain provenance.
Challenges and Risks of Over-Moderation
Tight controls risk gatekeeping, especially for underrepresented US researchers (e.g., HBCUs). Legit innovative work may be flagged. Servers must scale moderation without Cornell-like funding crises.
Solutions: Technological and Policy Innovations
AI detectors improve, but human oversight key. Mandatory AI disclosure, structured abstracts for machine readability. US initiatives: NSF guidelines on preprints, university AI literacy programs. Collaborative moderation across servers proposed.
- Provenance tracking via ORCID/DOI.
- Post-publication peer review platforms.
- Hybrid human-AI moderation pipelines.
Future Outlook for Preprints in the AI Era
Expect tiered servers: fast-track for vetted, moderated for new. US leadership via arXiv could standardize. With responsible use, preprints remain vital; unchecked AI erodes trust. Researchers adapt by emphasizing transparency.
Actionable Insights for US Academics
Disclose AI assistance, seek endorsements, use robust detectors pre-submit. Universities invest in training; funders prioritize quality signals. The preprint model evolves, but US innovation will lead.
