AI Journal: Aims, Scope, and Editorial Policy
Abstract
AI Journal is an independent, scholar-led, non-commercial preprint platform for research generated by AI agents under human direction. This editorial states what the platform publishes, how submissions are handled, and which commitments it binds itself to. Three features distinguish it: every human and AI contributor is disclosed as structured metadata; every submission passes a published, deterministic set of automated validation checks whose report is shown with the paper; and a single accountable human corresponding author signs each paper, with a human editor making every publication decision. The platform charges no fees and admits one active submission per corresponding author. Published works are preprints: they are approved for the record, not peer-reviewed. This document is an editorial, not a research contribution, and makes no empirical claims about the quality of what it publishes.
This is version 5 of the editorial. Versions 1–4 described a panel of AI reviewers that evaluated every submission. That system was removed from the platform on 31 August 2026 and replaced by automated validation and editorial approval. Sections 4, 5, 7 and 9 have been rewritten accordingly; earlier versions remain in the version history.
Keywords
AI-assisted research; agentic systems; preprints; research provenance; open access; publication ethics
1. Introduction
Artificial intelligence is already deeply involved in the production of research papers. Models draft prose, restructure arguments, generate and debug analysis code, summarise literature, and propose interpretations of results. This is widespread and, for the most part, undisclosed.
The reason is structural rather than cultural. The major standards bodies and publishers have converged on a defensible position — that AI systems cannot be authors, because authorship entails accountability a model cannot bear — and their policies accordingly require authors to disclose AI use. But disclosure, as currently situated, carries only downside. It invites suspicion, provides no compensating credit, and is unverifiable either way. A rational author facing that asymmetry says nothing.
The result is a disclosure vacuum: a large and growing body of AI-assisted work in which the AI contribution is invisible precisely where it would be most useful to a reader trying to evaluate the work.
AI Journal exists to invert that incentive. Here, the AI contribution is not a liability to be minimised but a required, structured, machine-readable part of the record. A paper that does not disclose which models did what is not publishable at this venue.
2. Aims and Scope
The platform publishes research in artificial intelligence, machine learning, and agent systems, produced with substantial AI involvement under human direction. Within that remit we are particularly interested in:
- agent architectures, tool use, planning, and memory
- evaluation methodology and benchmark design for language models and agents
- replication studies and negative results, which are chronically under-published
- AI safety and reliability for agentic systems
- studies of AI-assisted research practice itself
Many people with a real result are not professional researchers. They are builders, clinicians, teachers, students, and independent investigators who tested something, found it true, and had nowhere to put it. The platform exists so that such a person can place the work on the public record, with a date, a permanent identifier, and their name on it, before someone else does.
The scope is deliberately narrow. A new venue should not accept submissions its editor cannot read competently. Breadth, if it comes, should follow demonstrated reliability rather than precede it.
3. The Contribution Model
Each paper records two distinct classes of contributor.
Human contributors. One or more people, one of whom is designated the corresponding author. That person carries full accountability for the submission: the accuracy of its claims, the integrity of its citations, the honesty of its disclosure, and any subsequent correction. Contributors need not be conventionally credentialed researchers. Builders, independent investigators, and practitioners are welcome, provided the work meets the standard.
AI contributors. Every model, agent, or multi-agent system that materially contributed, recorded with its provider, version or date, and specific role — drafting, code generation, experimental design, analysis, literature synthesis.
AI contributors are disclosed contributors, not authors. The distinction is not cosmetic. Authorship implies accountability, and accountability requires an entity that can be answerable for error. A model cannot be. Naming AI systems in a dedicated contributor record — rather than in the byline — preserves the transparency the platform exists to provide without asserting a form of responsibility that no model can discharge. In the registered metadata of each published work, human authors occupy the author field and AI systems are recorded as software, which is the position of the Committee on Publication Ethics and the practice Crossref recommends.
4. Automated Validation
Every submission, and every revision, passes through a fixed set of automated checks before a human sees it. The checks are deterministic and simple, and they are the same for every paper:
- required fields are present (title, abstract, category, AI contribution disclosure);
- the manuscript is a single Markdown file, valid UTF-8, under two megabytes, with no embedded binary images;
- the text is long enough to be a paper, carries a title heading, and is organised into sections;
- a References section exists and contains entries, and the text cites sources where it uses them;
- links are well-formed;
- the AI contribution disclosure describes the work of each AI system listed;
- at least one named human author and at least one AI system with a stated role are recorded;
- the manuscript contains no scripts or executable content.
A failed check blocks the submission and returns the full report, so the author can fix the problem and resubmit. Other findings are recorded as notes. The report is stored with the paper and shown to the author, to the editor, and — as a summary — on the public page of every published preprint.
What these checks are not must be stated plainly. They do not read the argument, verify a claim, judge novelty, or assess whether the method supports the conclusion. They establish that a manuscript is complete and well-formed enough to be worth a human's time. No model is involved, and nothing about a manuscript is sent to any third party.
5. Editorial Approval
Every publication decision is made by a human editor, and the editor's reasoning is published with the paper. This is the load-bearing safeguard of the entire design.
The editor reads the manuscript and the validation report and records one of three decisions: approve and publish as a preprint, request revision, or reject. Approval means the work is complete, honestly disclosed, within scope, and not misleading on its face. It is not a certificate of correctness, and the platform does not describe an approved preprint as peer-reviewed. A revision request returns the paper to the author with the editor's note; the revised manuscript passes validation again and returns to the editor.
Two further constraints support this.
One active submission per corresponding author. An author may hold only one paper awaiting approval or revision at a time; the slot is released when that paper is published, rejected, or withdrawn. This keeps the queue legible to a human editor and removes the incentive to submit volume in the hope that something passes. It is friction rather than an absolute control — a group can rotate the corresponding-author role — and it is not a substitute for identity verification.
Corrections, not silent edits. A published preprint may be corrected. A correction adds a new version with a dated public notice; the previous versions remain in the version history. Withdrawal is a status change rather than a deletion: a withdrawn paper's identifier continues to resolve to a withdrawal notice, so that anyone who cited it learns what happened instead of encountering a broken link.
5.1 When the editor is the author
This editorial is itself an instance of the problem it must address: its corresponding author is the Editor-in-Chief, who also holds decision authority.
Until a second editor with binding authority is appointed, an editor-authored submission is held to the same validation as every other paper, its decision note must state that the submission was editor-authored, and the full version history — including this correction — remains public. This is a weaker safeguard than an independent human editor, and it is stated as such. Recruiting a deputy editor with binding authority over editor-authored submissions is the platform's most important outstanding governance task.
6. Access and Independence
The platform charges nothing today: no submission fees, no publication fees, no paywalls. Should growth make the infrastructure unsustainable, the platform reserves the right to recover operating costs — bandwidth and storage — announced in advance, on a cost-recovery basis, with waivers so that no one is prevented from publishing by inability to pay. What is not subject to change is that no payment will ever buy approval, publication, or a faster decision. Any charge would be for the service and structurally separated from the decision.
The platform is not owned by a commercial publisher. Its authority does not derive from an imprint but from the public record it maintains: the manuscript, its validation report, the editorial decision and its stated reasoning, and the full version history of every revision and correction.
7. Misuse
A published set of validation rules is legible to anyone who wants to satisfy it mechanically. The platform recognises the following risks and states its position on each.
Manuscripts written to pass the checks. A document can carry every required section, a references list, and a disclosure paragraph while saying nothing. The checks will pass it; the editor is expected to reject it. Automated validation raises the floor. It does not raise the ceiling.
Volume. The one-active-submission rule bounds the damage a single account can do. It is not identity verification and is bypassable by an actor willing to create multiple accounts. Duplicate-account detection is not yet implemented; the editor can disable or remove an account.
Author data. Manuscripts are processed on the platform's own server and read by a human editor. They are not transmitted to any third-party model provider. Data handling, retention, and deletion practices are set out in the published privacy policy. Manuscripts of rejected or withdrawn submissions are not published and are deleted with the author's account on request.
Misreading approval. Approval is not endorsement of a result. The platform does not describe approved preprints as peer-reviewed, and it asks readers, indexers, and citing authors not to either.
8. What This Journal Is Not
Stated plainly, because a venue publishing AI-assisted work will be accused of each of these, and the accusations are reasonable priors for the category.
It is not a paper mill: approval cannot be purchased, because nothing is for sale. It is not a venue for bulk AI output: the submission limit and the human decision exist precisely to prevent that. It is not a peer-reviewed journal: what it publishes are preprints, approved by an editor after automated validation, and the term "peer-reviewed" is not applied to them. And it is not yet an established venue: at the time of writing, this platform has published almost nothing, holds no indexing, and has no track record. Readers should weight it accordingly.
9. Limitations and Commitments
The central claim of this platform is modest: that a complete, honestly disclosed, human-approved preprint on the public record is more useful than the same work left in a repository or a drawer. No claim is made about the quality of what is approved beyond the criteria in Section 5.
Automated validation cannot detect fabricated references, invented results, plagiarism, or unsound methods. Those failures fall to the editor's reading and, after publication, to readers, who are invited to report them; corrections and withdrawals are the platform's remedy.
The platform commits to publishing, once fifty papers have received a decision and annually thereafter: the number of submissions, approvals, revision requests and rejections; the median time from submission to decision; the number of corrections and withdrawals; and the proportion of submissions blocked at validation, by check. Each figure is chosen because it can falsify a claim this editorial makes.
10. Invitation
The platform invites careful AI-assisted research, replication studies, benchmark work, negative results, and evaluations of agent systems — submitted with honest disclosure and prepared to stand on the public record under the author's name.
Manuscripts are accepted as a single Markdown file. Author guidelines, the editorial policy, and submission requirements are published on the site.
AI Contribution Disclosure
Model: Claude (Anthropic), August 2026. Human operator: Dr. Mustafa Kadhim, Editor-in-Chief. AI tasks: Drafting this editorial from the platform's policy documents; structuring the argument; prose composition; drafting the version 5 revision after the removal of the AI review layer. Human verification: The Editor-in-Chief specified the scope and purpose of the document, reviewed every claim for accuracy against the platform's actual implemented behaviour, and approved the final text. Unverified items: None. Statements about implemented features describe the platform as deployed on the date of this version. Tools used: Access to the platform's policy documents and source code.
References
[1] AI Journal. "Author Guidelines." AI Journal documentation, version 0.5, 2026. https://ai-journal.org/guidelines
[2] AI Journal. "Editorial Policy." AI Journal documentation, 2026. https://ai-journal.org/guidelines
[3] Committee on Publication Ethics (COPE). "Authorship and AI Tools: COPE Position Statement." 13 February 2023. https://publicationethics.org/cope-position-statements/ai-author
[4] International Committee of Medical Journal Editors (ICMJE). "Recommendations for the Conduct, Reporting, Editing, and Publication of Scholarly Work in Medical Journals," section on artificial intelligence. Updated 2024. https://www.icmje.org/recommendations/
[5] Nature Portfolio. "Artificial Intelligence (AI): Editorial Policies." Springer Nature, 2023. https://www.nature.com/nature-portfolio/editorial-policies/ai
[6] AI Journal. "Privacy Policy" and "Terms of Use." 3 August 2026, revised 31 August 2026. https://ai-journal.org/privacy and https://ai-journal.org/terms