How should science be done in the age of AI?

XSci is a preprint platform where a paper's reviews are published alongside it, free to read, with the authors' replies on the same page. Our mission is to make the validation of science as open, as fast, and as fairly rewarded as its production. As far as we know, no other venue lets the world watch AI review being tested against real papers in public.

The problem we are built for

Science has never produced so much, and never checked so little of it.

More papers are written today than at any point in history, and AI is accelerating the writing faster than anything has ever accelerated the reading. Meanwhile the machinery that checks this output is essentially the machinery of 1950: a small number of unpaid experts, recruited slowly by editors, writing private letters that most readers will never see. Review takes months. The reports are thrown away or locked in editorial inboxes. The most informative discussion about a paper, what a careful reader found wrong with it and how the authors answered, is treated as waste.

The bottleneck of science has moved. It is no longer hard to produce a paper. It is hard to know which papers to trust. That gap between production and validation is widening every year, and the tools that widened it are not going away. They are moving deeper in: AI now proposes hypotheses, runs analyses, and in some fields drafts entire papers. A literature partly produced by machines cannot be checked by unaided human attention.

We think the answer is not to slow the production down. It is to bring the validation up to speed, and to do it in the open.

What is new here

Many researchers already use AI to judge papers. They do it privately. What nobody does is publish many named AI review systems side by side on the same real papers, with paid expert judgement on top of them and the authors answering back on the same page in public. That is the new thing here, and the rest of this document follows from it.

Four things make it work.

One place, and more than one model. AI review services exist, and every one of them sells you its own model. We put several providers' systems in front of you and let you choose between them. You run the systems you want from a growing set of integrated providers, in one workflow, and each report carries the name of the system that produced it. We want you to run them and we want you to publish what they say, so publishing the reports costs you less than keeping them private.

Expert judgement is worth paying for. Reviewing a paper properly takes hours of the scarcest attention in science, and the people who do it best are paid nothing for it. We think that time and expertise have a price, and we pay it. Payment is also what lets us solve the two problems goodwill cannot: matching a paper to someone who actually knows its methods, and giving that person a reason to meet a deadline. Science gets checked faster here because someone is accountable for delivering the check.

We do not grade the systems we host in private. Our research team evaluates every AI system independently before it appears here and keeps evaluating it afterwards. Their strengths and their failures then sit side by side on the same papers, in public, where anyone can compare them. We can afford to be honest about which systems perform well because we are not betting the company on any one of them.

Two layers, and you pick the ones you want. AI reports and expert review are separate services. You can run AI reviewers and stop there, or go straight to a paid expert without running any AI reviewer. Where you buy both, the reports are on the page before the expert opens the paper, so none of their paid hours go on ground the reports have covered. What is left is judgement, and a written standard to exercise it against. The standard is simple enough to state in full: certification means a qualified expert judges that the paper, after revision, has the potential to be published in the journals or conferences you are aiming at, or, where you name none, in the leading venues of its field. Disciplines differ too much for one rubric to mean the same thing everywhere; a shared expectation does, and a certificate whose meaning you cannot look up is a logo. This arrangement is why certification here can be quick without being cheap.

What this does for you, as an author

This is the whole service, in the order you meet it.

Free, immediate publication. You post your manuscript, in any discipline, and it is public at once with a permanent identifier, before any review has run. Your priority is established the day you deposit, not the day someone else's process finishes.

Your choice of AI reviewers, in the open. You choose whether to run AI reviewers at all, and which ones, from a growing set of integrated providers, in one place and one workflow. The reports arrive within hours and can be published alongside your paper. They are labelled for what they are: reports, produced by named systems. They are never called peer review, and no machine output certifies anything.

Nothing goes public without you. You see every report before the world does, and you decide, report by report, what is published. Openness costs less here, because it is what we exist to build; it is a choice we price, never a condition we impose.

Expert certification, when you want it. You can request review by paid domain experts at any point, whether or not you have run a single AI reviewer. Certification is reserved for this: it means a qualified human has read the paper and it met a published standard. Expert reviews can be published with the paper; reviewer identities stay protected. Certification is optional. It is a distinction your paper can earn, never a gate it must pass to exist.

Experts chosen in the open. How we find reviewers is not a secret. For each paper we invite experts from three groups: the authors of work the paper directly cites, the close collaborators of those authors, and researchers who have published on similar questions in the journals or conferences you are aiming at, where you name them. So you know, before anything begins, what kind of reader your paper will meet. Few venues will tell you that much about where their reviewers come from.

You answer back. Every review, human or machine, can be publicly answered by you at no cost. When a report is wrong, the strongest response is not to hide it but to say, in public, why it is wrong. That exchange is exactly the open discussion the platform exists to host.

The record travels with the paper. The reports, any expert assessment and your replies to both stay attached to your work. An editor deciding whether to take the paper further can read all of it before commissioning anything, and so can whoever funded it. You are not asking anyone to take the quality of your work on trust, and you are not starting from nothing when the paper moves on.

Reading requires nothing. No account, no institutional login, no paywall, no request for your email. That includes the reviews.

One consequence of these commitments is our pricing. Openness is subsidised: authors who publish their review reports pay less, in many cases nothing, because public review is the asset we exist to build, and we would rather fund it than advertise. Authors who prefer to keep reports private pay the full cost of the service. The discount is not a promotion. It is the mission, expressed as a price.

What this does for you, as a reviewer

We pay you. Your time and your expertise have a price, and we would rather pay it than ask for your goodwill a fourth time this month. Payment buys your time and never your verdict: a judgement stands as written, whatever it says. Reviews are single-blind: your identity is not disclosed.

You are also matched to papers whose methods you actually know, because that matching is a service we are paid to get right rather than a favour an editor asks around for.

And you work with the best tools we can build. Our reviewers use AI, and we say so. Where the author has run AI reviewers, those reports are on the page before you open it, so what remains is the part machines cannot do, which is judgement. We see no reason to stigmatise the use of AI in review while the rest of science uses it every day. The line that matters is not whether a tool was involved but who is accountable for the conclusion. Here the expert makes the final call and answers for what they sign, and our job is to hand you better instruments than any journal ever has.

What we believe in

Review belongs in the literature, not in an inbox. A paper's reviews are part of the scientific record. On XSci, the paper, the AI reports run on it, the expert assessments, and the authors' replies are published together on one page, readable by anyone, with no account required. A reader arriving at a preprint should not face it alone; they should see what has already been checked, what has been contested, and what the authors said in response.

The work, not the byline. Posting on XSci is free and open to any author willing to stand behind their own work. No paper is more visible here because someone paid for it: there is no promoted placement and no ranking money can buy. A paper's standing here comes from the open record attached to it: what the reports found, what reviewers said, how the authors answered. It does not come from the prestige of a venue or the affiliation of an author. We will not pretend this ideal is fully realised on day one, but every design decision we make points at it.

A disagreement is a finding. Independent reviewers will disagree, and we do not smooth that over. Where several systems raise the same point, the convergence is information. Where they conflict, the conflict is itself a finding, and it is published like everything else. We would rather show an honest disagreement than manufacture a consensus nobody stands behind. We are building the place where AI review is tested against real papers, at the scale of the literature.

What we will not sell

Three things are routinely for sale in this industry. Refusing to sell them costs us money, and we have chosen to pay it.

Publication charges pay a venue to say yes. Open-access publishing largely runs on a fee the author pays on acceptance, which means a venue earns more the more it accepts. That incentive has damaged publishers far larger than us, and researchers have read the results. We charge nothing to publish, and nothing can make a paper more visible here, so we have walked away from the largest revenue stream in open-access publishing. Our revenue comes from review services instead.

A machine verdict is cheap, and worth what it costs. Expert time is the single largest cost in our model, and dropping the human pass would be the easiest margin we could ever take. It would also mean selling you a certificate that nobody read your paper to issue. We priced that temptation out: certification means a qualified human read the paper, or it does not exist.

Every AI tool in this market tells you its own model is the best one. None of them can show you, because none of them publishes its model's failures next to a competitor's on the same paper. We integrate other people's reviewers and name them on every report, which costs us the proprietary-AI story and gives you an honest comparison in its place.

Where this goes

Open access gave everyone the paper. It did not give anyone a reason to trust it. The next step for open science is a literature where every paper carries its own public record of scrutiny: what was checked, by whom or by what, what was found, and what the authors said back.

That step matters more because of what is coming. Some of the scientific frontier will soon be moved by machines, and their output will need checking no less than ours does. A validation layer fast enough for machine-scale production will itself use machines. One trustworthy enough to matter must work in the open. Speed requires machines; trust requires openness. This platform is built on the wager that the two together can bring the checking of science back into balance with the making of it.

That record, accumulating paper by paper, is what we are building. Not a gatekeeper, and not another tool, but the open validation layer science is missing.