According to science.org, a new open platform called aiXiv is fully embracing artificial intelligence by accepting both AI- and human-authored work and using built-in AI reviewers for quality screening. Founded by Guowei Huang, a Ph.D. candidate at the University of Manchester, and collaborators from institutions like the University of Toronto and Oxford, the platform launched after a mid-November update and currently hosts just a few dozen papers. The system uses five AI “agents” to assess submissions for novelty and soundness, posting work if three recommend acceptance, and it can generate reviews in 1-2 minutes. This comes as traditional preprint servers like arXiv are banning certain AI-written documents, and others, like bioRxiv’s operator, are adding AI review tools to handle surging submissions. AiXiv’s team claims preliminary tests show their iterative review loops improve AI paper quality, and they have unpublished data comparing their AI reviews favorably to human ones from an October robotics conference.
The AI Review Experiment
So here’s the thing: aiXiv isn’t just letting AI write papers—it’s putting AI in the editor’s chair. The process is basically a fully automated loop. You submit, five LLM agents tear it apart (or praise it), and you get feedback in minutes, not months. You can then revise and resubmit, creating this iterative, machine-driven polishing cycle. The founders argue this is a necessary infrastructure for the “AI-era of scalable scientific output.” And look, they have a point about speed. When the human peer-review system is buckling under a tsunami of papers—many now AI-assisted—a 2-minute review sounds like a miracle cure.
But the big question is, what exactly are these AI reviewers evaluating? They’re phenomenal mimics, as Thomas Dietterich from Oregon State University points out. They can catch a mismatched reference or a numerical inconsistency in a heartbeat. But can they tell if the research is real? Or if the core idea is truly novel versus just a slick recombination of existing tropes? Early experiments, like the Agents4Science conference, show these systems often give overly generous scores on novelty and impact. They’re grading the packaging, not the substance. And on an open preprint server with “a lot more heterogeneity and slop,” as Stanford’s James Zou notes, that’s a dangerous gap.
The Transparency Gamble
One of aiXiv’s core arguments is actually pretty compelling. Most journals and servers ban listing AI as an author. Guowei Huang calls that lack of transparency “totally unacceptable,” and he’s right. That policy basically incentivizes researchers to use AI covertly. By being upfront about AI involvement, aiXiv is forcing the issue into the open. It’s saying, “We know this is happening, so let’s build a system that accounts for it.” This pushes the conversation toward what Carnegie Mellon’s Nihar Shah mentions: the need to verifiably track provenance, whether the work comes from a human, an AI, or some collaboration.
But there’s a massive risk, flagged by bioethicist Sebastian Porsdam Mann. If this platform becomes a “dumping ground” for low-quality, AI-pumped volume, it could delegitimize the entire idea of AI-led inquiry. It could taint even the good work. The defense mechanisms, like detecting hidden prompts in manuscripts, are clever, but the arms race between AI authors and AI reviewers is just beginning. Who wins when both sides keep getting smarter?
The Bigger Picture for Science
In a way, AI is just holding a mirror up to science’s existing problems. As Porsdam Mann says, a lot of the issues with AI review—superficial feedback, fraudulent research, overwhelmed systems—are already true of the current human system. AI just amplifies them and makes them impossible to ignore. AiXiv’s experiment, whether it flies or crashes, is accelerating a crucial discussion we can’t afford to delay.
Is the future one where AI generates thousands of plausible-but-shallow papers, and other AI systems sift through them? Maybe. But the optimistic take from aiXiv’s preprint and their unpublished data is that iterative AI review might actually improve the rigor of AI-generated work. It’s a fascinating, high-stakes beta test. The platform itself is at aixiv.science, and it’s worth watching. Because Nihar Shah is probably right: the “almost inevitable outcome” is AI doing a lot of research. The question isn’t *if* but *how* we build the systems to manage that future. AiXiv is one of the first real attempts to provide an answer.
