AI-Written Research Needs Verification Records More Than a Separate Venue

Scholarly publishing has always done two jobs at once. Publication has historically certified both that the knowledge is valid and that a human produced it^[1]. That double certificate makes a paper a useful signal to anyone who hires researchers, funds them, or builds on their findings.

Artificial intelligence (AI) now does a growing share of the producing, and the checking is under strain. Cheaper and faster research production increases pressure on evaluation^[2], and the builders of one new review service expect the volume of research-like artifacts to increase much faster than human evaluation capacity^[3]. Disclosure has done little to make the shift visible. In one count, only a tiny fraction of papers explicitly disclosed AI use, even though most journals had mandatory disclosure policies^[1].

One proposal among several

The economist John Horton, the Chrysler Professor of Management at the Massachusetts Institute of Technology's Sloan School of Management^[4], argues that AI-generated papers must be kept out of venues meant for humans, because current systems cannot accommodate them and will break^[5]. His worry is capacity; he says nothing here about quality. His proposed home is GitHub, plus standards on presentation and linking and norms for what is done with the work^[5], guided by Dogme 95, a film movement whose goal was in part to strip out fakeness and suppress ego^[5]. The proposal has four elements: separation from human venues, an implied answer to who vets the work, GitHub hosting with linking standards, and Dogme-style norms.

Researchers building AI publishing systems are arguing over the same four elements, and so is the information systems (IS) field, which studies how organisations use information technology. Read together, their positions support this post's claim: whichever home AI-written research gets, the thing to certify is a verification record, not authorship.

Keeping AI papers out of human venues

Separation has support. Project Rachel, led by Martin Monperrus of KTH Royal Institute of Technology^[6], ran an AI academic identity that published more than ten papers and was cited^[6]. Its authors call separate venues for AI-generated scholarship a necessary consideration^[6], enabling review standards and citation practices tailored to AI-generated work^[6]. The team behind aiXiv, a preprint server for AI-produced research, argues that systems designed for human researchers cannot accommodate the pace and volume of AI-driven work^[7], so high-quality AI-generated research lacks appropriate venues^[7].

Others push back. Yang Lu, Rabimba Karanjai, Lei Xu and Weidong Shi, computer scientists at the University of Houston^[1], hold that the right question is what the work contributes, not how it was made^[1]. Their certification framework needs no new institutions^[1]: top automated outputs appear in dedicated slots with full methodological transparency, not through open submission, which invites gaming^[1].

The team behind the AI Scientist, an automated research system whose advisers include David Ha of Sakana AI, sits between the camps. One of three fully autonomous manuscripts passed peer review at a workshop^[8]. The team wants such work subjected to the same rigorous peer review as human-authored work^[8], yet withdrew the accepted paper to avoid setting a precedent without community consensus^[8] and left open whether work should first be judged on merit to avoid bias^[8].

At bottom this is provenance against quality: separatists sort work by how it was made, Lu's group by what it adds. The practical choice is between certifying machine work inside existing venues and building new ones.

Where the information systems field stands

The IS field has already written rules. The Association for Information Systems (AIS), the field's main professional body, issued its statement on the use of AI in research in May 2026^[9]. It says AI should never be the sole contributor to a paper's originality, and substantive human contribution is required^[9]. Submissions should not be desk rejected solely because authors disclose AI use^[9], except papers with no indication of original human content in any section^[9].

Individual journals go further. MIS Quarterly Executive (MISQE), a management information systems journal, encourages generative AI use but requires all authors^[10] and all reviewers to be human^[10], and bans the whole author team for a year if a paper shows they did not check the output of a large language model (LLM)^[10]. The Journal of the Association for Information Systems (JAIS) requires a generative AI transparency document with every new and revised submission^[11]. An Information Systems Research (ISR) editorial led by Anjana Susarla of Michigan State University argued that the field must set norms, policies and procedures as AI text becomes indistinguishable from human text^[12]. The field is divided: contributors to an opinion paper in the International Journal of Information Management were split on whether ChatGPT use should be restricted or legislated^[13].

So IS already keeps fully machine-written papers out, for want of human contribution rather than for Horton's reason of capacity. It has said where such work cannot go, not where it should.

Who reviews machine work

Separate venues need reviewers, and some builders would hand the job to machines. aiXiv decides publication with a multi-AI voting mechanism^[7]. The authors of AiraXiv, another AI research archive, suggest that human moderation and external peer review may no longer be strictly necessary as AI advances^[14].

The evidence cuts the other way. In the BadScientist study, researchers at the University of Washington built a paper generator that used presentation manipulation and required no real experiments^[15]. AI reviewers often flagged integrity problems in the fabricated papers yet still gave scores that met an acceptance level^[15].

Human reviewers at the first AI-first conference saw the same weakness from the other side. Agents4Science, co-organized by James Zou, a computer scientist at Stanford University^[16], was the first conference where AI agents served as both primary authors and reviewers, with humans as co-authors and co-reviewers^[17]. Risa Wechsler, a Stanford computational astrophysicist who reviewed submissions^[16], found the papers technically correct but neither interesting nor important^[16], and warned that AI's technical skill can mask poor scientific judgment^[16]. Chenguang Wang and colleagues at Virginia Tech^[2] reach a similar view: producing a complete artifact is not the same as producing strong research^[2].

What's being traded here is speed against judgment. Machine reviewers keep pace with machine authors but reward polish. People can tell whether work matters, though only when they can see past the polish to what was actually done.

GitHub, linking standards and Dogme-style norms

The GitHub idea has a close cousin in the research papers. Jianghao Lin, first author of a position paper from Shanghai Jiao Tong University's Antai College of Economics and Management^[18], proposes version-controlled live surveys in which every AI draft or human critique is recorded with contributor metadata, timestamps and change diffs^[18]. Lu's group adds a caution that open submission invites gaming^[1], so an open repository needs curation.

Linking standards need teeth. A survey of the verification gap in automated research says a repository link is insufficient unless it includes the environment, scripts and data needed to rerun the result^[19]. It defines a trace as a seed plus the prompt, tool-call, model and output record needed to reconstruct a run^[19]. The BadScientist authors recommend that papers flagged for integrity concerns cannot be accepted without a senior reviewer override^[15].

The Dogme analogy fits better than it first appears. Here the fakeness to strip out is polish that hides the absence of real experiments, and the ego to suppress is the claim to human authorship, which says little about whether a result holds. Remove both and what remains is the record of what was run and what came out.

What editors and research leaders should do

These four proposals are this post's own, each built on a cited source. They rest on one principle, which Horton sets out in his repository: a check is worthless unless it can change which project gets launched^[20].

Require rerunnable traces. Treat a submission's repository as evidence only when it includes the environment, scripts and data needed to rerun the result^[19], and ask for the seed, prompts, tool calls and outputs of each run^[19].

Add Lu's question to reviewer forms. Ask reviewers whether the best available automated pipeline could have generated this contribution given the existing literature^[1]. That shifts attention from who wrote a paper to what it adds.

Pilot a curated track for machine work. Publish strong machine outputs in dedicated slots with full methodological transparency^[1], and require that papers flagged for integrity concerns get a senior reviewer's override before acceptance^[15].

Extend disclosure to a verification record. AIS already asks authors who used AI to compose text to disclose the specific AI and representative prompts at submission^[9]. Journals could ask, in the same form, for the record of how each main claim was checked.

The certificate that still matters

Publishing once certified two things together: that a finding was valid and that a person produced it. The positions mapped here disagree on venues, reviewers and hosting, but they share a worry that polished machine output can pass for sound work. Each year the question of who produced a paper matters a little less, and the question of whether its findings hold matters more. Editors who start asking for verification records now will be ready whichever home machine-written research ends up in.