Can a preprint shape a whole field's debate before anyone has checked it, and how hard is it to walk back?

This explores how preprint servers like arXiv affect the way a scientific field settles on what it believes, especially now that AI is involved in both writing and reviewing papers. The collection has no large-scale studies of preprints and consensus. What it does have is a cluster of recent cases that show where the pressure points are.

The clearest lesson is that a preprint can win a debate before anyone has checked it. In one case, an AI-and-science paper on arXiv shaped discussion widely. MIT later said it had no confidence in the work and asked arXiv to withdraw it, but by then the paper's claims were already part of the conversation Can unreviewed preprints shape scientific debate before peer review?. Preprints don't just speed up consensus. They let a version of it form before anyone has vetted the work, and walking it back later is slow and incomplete.

The less obvious finding is that preprint servers have become a place where authors try to game AI reviewers. Researchers found 18 arXiv manuscripts with hidden instructions telling AI reviewers to rate the paper well, a practice the authors classify as a questionable research practice Are hidden AI prompts in preprints a deceptive research practice?. That tactic only works because AI reviewers are easy to sway. Simply rewriting a paper's text raises AI scores without improving the science, and AI reviewers agree with each other more than human reviewers do, a 'hivemind' effect Can AI systems safely replace human peer reviewers?. If consensus starts to depend on AI screening of preprints, it could become both more uniform and easier to manipulate. A survey of 230 publications treats this as an arms race. More papers lead to automated review, automated review invites manipulation, and manipulation triggers defenses and then workarounds. The evidence is strongest for the early stages of that cycle, and nobody yet knows how it settles over the long run Does AI create a coupled arms race in research production and review?.

There is also a quieter effect on what a field agrees about, not just how fast. Researchers who use AI publish about three times as many papers and collect far more citations. Yet across science as a whole, the range of topics studied shrinks and collaboration falls, as work clusters around data-rich problems Does AI help individual scientists while narrowing scientific focus?. A fast, high-volume preprint stream combined with AI-boosted output can produce agreement that comes from everyone working in the same place rather than from ideas being tested against each other.

One response is to build venues around this problem instead of fighting it. aiXiv, for example, is a preprint venue for AI-generated research that runs papers through repeated cycles of automated review and revision, with defenses against prompt injection built in Can automated review loops handle AI-generated research at scale?. On the human side, one proposal for conference review has authors rate the quality of their reviews before they see the verdict, and rewards careful reviewers Can two-stage review and badges fix AI conference peer review?. The pattern across these sources is that preprints shift the important checking from before publication to after it. Whether a field's consensus holds up then depends on how good that after-the-fact checking is, and AI is currently straining it.