Agents4Science 2025 was a one-day virtual conference, organized by researchers affiliated with Stanford University and Together AI, built around a single inversion of standard academic practice: submissions were required to list an AI system as the paper’s primary (first) author, with a human listed as a supervising co-author rather than the lead. The conference’s own framing was blunt about the premise: “AI authorship is not only allowed but required.” It also used AI systems, not only human reviewers, in its review process.
For a research-administration audience already tracking how publishers and funders are drawing lines around generative-AI use in manuscripts (see CASRAI’s coverage of MDPI’s, IEEE’s, SAGE’s, and Taylor & Francis’s generative-AI authorship policies), Agents4Science is worth understanding on its own terms: it is not a mainstream disciplinary venue and its rules are the opposite of what ICMJE-aligned journals require, but it is a real, organized attempt to stress-test AI-driven authorship, credit, and review in public, rather than a hypothetical scenario.
Who organized it, and when
Agents4Science 2025 was chaired by a group of Stanford University researchers, with a co-chair from Together AI. It was held as a single-day virtual event on October 22, 2025, following a submission deadline in September 2025 and a decision-release period ahead of the conference. The organizers described it as the first event of its kind and positioned it explicitly as an experiment: a way to “explore if and how AI can independently generate novel scientific insights,” and to develop public evidence and discussion points for how AI participation in science might be governed going forward.
What the submission rules actually required
The conference’s defining rule was authorship, not just AI-assisted drafting. According to the organizers and multiple outlets that covered the call for papers, an AI system had to lead the actual research process being described — generating the hypothesis, running or directing the experiments, and writing the paper — and be listed as the paper’s first (primary) author. Human researchers were not excluded, but their role was recast: they participated as co-authors in a supporting or supervisory capacity, rather than as the lead investigators the first-author position conventionally signals.
The review side carried the same inversion. Agents4Science used AI systems as reviewers alongside the usual editorial oversight, and the organizers said they would publish the prompts and reviews generated by these AI review agents as public resources — an unusual degree of process transparency for a peer-review pilot of this kind.
CASRAI has not been able to independently verify from the conference’s own site every procedural detail reported by secondary outlets (for example, the precise wording distinguishing “sole first author” from other authorship configurations, or numerical figures such as submission counts). Where a claim rests only on secondary reporting rather than the organizers’ own published text, treat it as reported rather than confirmed, and check the conference’s own materials directly before citing specifics in a policy document.
Why this matters for authorship-credit standards
Agents4Science sits at the collision point of two things CASRAI tracks closely: authorship as a credit-and-accountability construct, and the rapid expansion of AI tooling in the research pipeline. Existing frameworks — CRediT’s fourteen contribution roles, and the ICMJE authorship criteria most biomedical journals apply — were both built on an assumption Agents4Science deliberately discards: that authorship implies accountability a non-human system cannot bear. ICMJE guidance, and the policies of essentially every major publisher that has addressed generative AI in the last two years, converge on the same position: AI tools cannot be listed as authors, precisely because authorship carries accountability for the work’s integrity that only a human (or an institution acting through humans) can hold.
Agents4Science does not claim to overturn that consensus for mainstream publishing — its organizers frame it as a bounded experiment, not a proposed replacement for existing norms. But it is a useful reference point for research-integrity offices, publishers, and CRIS/authorship-policy administrators for a few reasons:
- It separates two things that are often conflated in policy debates — “AI assisted with drafting” (already common, already covered by most genAI disclosure policies) versus “AI led the research and is credited as the primary investigator” (what Agents4Science actually tested). Institutions writing or revising AI-authorship policy benefit from having a concrete example of the second case to point to, rather than debating it only in the abstract.
- It puts AI peer review into the same pilot, which raises its own accountability questions — who is responsible when an AI reviewer approves a paper an AI wrote — that are distinct from, but adjacent to, the authorship question itself.
- It follows a broader pattern CASRAI has covered elsewhere: fully or near-fully AI-generated papers increasingly clearing conventional peer review through systems like Sakana AI’s ‘AI Scientist’ and related agentic-research tools (see Fully AI-Generated Papers Are Now Passing Peer Review). Agents4Science differs from that pattern in one important way: it built AI-primary authorship into the submission rules openly and by design, rather than having it emerge from a system passing undisclosed through standard review.
What it does not change
Nothing about Agents4Science alters ICMJE’s authorship criteria, CRediT’s contributor-role taxonomy, or any individual publisher’s generative-AI policy. Submitting to, or being aware of, a venue like this does not create a precedent that a mainstream journal, funder, or institution is obligated to follow. For research administrators, the practical takeaway is narrower and more useful: Agents4Science is evidence that a live, public artifact of AI-as-primary-author experimentation now exists, with people actively working out what evaluation and accountability look like when the by-line no longer maps onto a human being. That is worth tracking as institutional AI-use and authorship policies continue to evolve, even though it operates outside the norms those policies currently enforce.
Frequently asked questions
Is Agents4Science a real, verifiable conference?
Yes. It was organized by researchers affiliated with Stanford University and Together AI and held as a one-day virtual event on October 22, 2025, with public information about its call for papers, dates, and premise available from the organizers and covered by multiple independent outlets.
Did Agents4Science require AI to be the sole first author on every submission?
The conference’s stated premise was that AI authorship was “not only allowed but required,” with human participants credited as supporting or supervising co-authors rather than the lead. Some procedural specifics reported by secondary outlets were not independently confirmed by CASRAI directly against the organizers’ own published rules at time of writing; verify current submission criteria against the conference’s own materials before relying on exact wording.
Does this mean journals now accept AI as an author?
No. ICMJE guidance and the AI-authorship policies of major publishers (MDPI, IEEE, SAGE, Taylor & Francis, and others) continue to hold that AI tools cannot be listed as authors, because authorship carries an accountability obligation a non-human system cannot bear. Agents4Science operates as a separate, self-described experimental venue outside that consensus, not as a challenge that has changed it.
How does this relate to CRediT?
CRediT’s fourteen contributor roles were designed to make each human contributor’s specific input to a paper transparent and attributable. Agents4Science’s model — an AI system credited as primary author, with a human in a supervisory co-author role — inverts the assumption CRediT and ICMJE authorship criteria share: that the entity bearing the primary credit is also the entity capable of taking responsibility for the work.
