Preprint Servers Tighten Moderation as AI Junk Science Floods Repositories

The moderation creep now visible across preprint infrastructure is a predictable second-order effect of two converging pressures: the commoditization of plausible-looking scientific text via LLMs, and the reputational damage preprint ecosystems absorbed during and after COVID-19 for hosting high-profile misinformation.

The core tension is architectural. Platforms like bioRxiv were explicitly designed as low-friction deposition systems — the value proposition was speed and openness, with the implicit contract that downstream readers (researchers, journalists) would apply their own critical filters. That contract has broken down. Citation of preprints by non-specialist audiences, combined with the near-zero marginal cost of generating structurally coherent but scientifically hollow manuscripts, has shifted the risk calculus for server operators.

Current moderation responses fall into three broad categories: (1) automated screening for AI-generated content using tools like GPTZero or in-house classifiers — unreliable at the margin and gameable; (2) expanded human triage, which scales poorly and introduces inconsistent editorial judgment; and (3) tiered visibility or "flagging" systems that preserve open deposition while surfacing risk signals to readers. The third approach is the most epistemically honest but requires UI investment and risks flag-blindness over time.

The prior art here is instructive. SSRN introduced basic screening years ago; arXiv has long used moderator networks with domain-specific expertise. Neither has solved the volume problem, and neither faced the current rate of AI-assisted submission. The question isn't whether some moderation is necessary — it clearly is — but whether the emerging frameworks will be transparent, consistently applied, and resistant to scope creep toward ideological or methodological gatekeeping.

Open questions worth tracking: Will moderation standards converge across servers, or fragment in ways that create arbitrage (bad-faith actors routing submissions to the least-screened platform)? Will AI-detection tooling improve fast enough to be operationally useful, or will it remain a liability that generates false positives against legitimate non-native English writers? And critically — who funds this? Moderation at scale is expensive, and most preprint servers run on thin institutional margins. If major funders (NIH, Wellcome, Gates) don't treat infrastructure support as a priority, the burden will fall unevenly, likely disadvantaging servers serving the Global South.