In March 2026, a scientific article generated by an automated system passed peer review and gained acceptance at an ICLR workshop. The total cost of the operation: less than fifteen dollars. The team at Sakana AI, in collaboration with the University of Oxford and the University of British Columbia, published this result in Nature. AI Scientist-v2 is thus the first system to produce a paper entirely generated by AI that has passed peer review evaluation at an ML conference workshop — even though the submission was made by the Sakana team, not by the system autonomously, and other systems claim similar milestones.
This fact, taken alone, says very little. What it implies for the infrastructure of knowledge says much more.
The Essentials
- AI Scientist v2, developed by Sakana AI with Oxford and UBC, is the first system to produce an article accepted in peer review without human intervention, for less than 15 dollars per paper, according to a study published in Nature in March 2026 and available on arXiv (arXiv:2504.08066).
- The system executes the complete cycle: hypothesis generation, design and execution of computational experiments, results analysis, manuscript drafting, submission and revision management.
- Current peer review is not designed to discriminate quality at this cost: the risk of an inflation of plausible but hollow publications is real and already documented for AI-assisted submissions.
- The central question is not technical. It is institutional: what guarantees, what standards, what governance allow this type of tool to accelerate discovery rather than dilute the literature?
Fifteen Years From Dream to Demonstration
To understand what AI Scientist v2 represents, one must recall where the field stood not long ago. In 2009, the Adam project — developed at Aberystwyth University in collaboration with Cambridge University — showed that a robot could formulate hypotheses in yeast biology and test them autonomously. It was an elegant demonstration, limited to a very circumscribed domain, and one that stopped at the laboratory door: the article remained the affair of human researchers.
The following decade saw two parallel mutations. Large language models acquired the capacity to produce coherent scientific prose. And automated execution environments — capable of launching scripts, interpreting outputs, modifying code in loops — matured to the point of being able to conduct small computational experiments without supervision. AI Scientist, in its first version published in 2024 by Sakana AI, assembled these two components. AI Scientist-v2 was developed in 2025 (April 2025 preprint); it was the human team at Sakana AI that submitted the papers generated by the AI to the ICLR workshop, not the system autonomously.
The result changes the nature of the problem. We are no longer talking about an assistant that helps a researcher write faster. We are talking about a system that poses a question, chooses a method, executes calculations, interprets results, drafts an article and manages exchanges with reviewers. The boundary between tool and researcher becomes blurred in a way that existing institutional frameworks did not anticipate.
What the System Can Do, and What It Does Not Touch
AI Scientist v2 functions in a specific domain: computational machine learning. It will not produce tomorrow an article on marine biology or Ottoman history. Its domain of choice is research where experimentation consists essentially of training and evaluating models on existing datasets, comparing architectures, measuring performance. It is a vast and active domain, but one where the boundary between experiment and calculation is thin.
What the system concretely does: it generates a hypothesis from a corpus of literature, proposes an experimentation protocol in code form, executes this code, interprets the outputs, draws conclusions from them and drafts these in standard academic format. Version 2 goes further by managing revisions requested by reviewers, adjusting the article and resubmitting. The article accepted at the ICLR Workshop represents genuine external validation, not an internal exercise.
What the system does not do, at least not yet: it does not propose experiments requiring new empirical data, does not touch the physical world, does not mobilize tacit knowledge or intuition about what deserves to be explored. It optimizes within a defined space. This is its strength and its limit. Its strength, because this space is precisely the one where the production of academic articles is most industrialized. Its limit, because great scientific breakthroughs rarely come from optimization within a known space.
The distinction between these two regimes of discovery — the routine of circumscribed exploration and the leap into the unknown — is at the heart of the debate. The former could be massified by systems like AI Scientist. The latter remain for now out of reach.
Measurable Acceleration, and Its Real Conditions
Proponents of scientific automation have a solid argument. Science suffers from a structural bottleneck: the number of questions that can be asked grows faster than human capacity to test them. In biology, computational chemistry, materials physics, thousands of plausible hypotheses await verification, not for lack of ideas, but for lack of hands and hours. A system capable of handling some of these at fifteen dollars per unit represents a real resource.
AlphaFold 2, developed by DeepMind and published in 2020, already illustrated this acceleration potential in a specific domain: protein structure prediction. In less than two years, it produced a database of more than 200 million structures, a task that would have taken decades using classical experimental methods. The result unblocked research programs in dozens of fields, from drug development to evolutionary biology. This is a case where automation clearly served discovery rather than replacing it.
AI Scientist operates on a different register — it does not predict, it explores and writes — but the logic of acceleration is similar. If the system reliably identifies combinations of parameters worthy of publication, and if its articles constitute a solid foundation for other work, the cumulative effect on the pace of research could be substantial.
The condition, precisely, is “reliably.” And this is where the institutional problem begins.
Peer Review Was Not Designed for This
The peer review system rests on an implicit assumption: each article submitted represents significant effort, and errors or shortcomings are detectable because they bear the traces of human thought. A researcher who gropes leaves marks. A system that optimizes text to pass a filter leaves different marks, and current evaluators are not equipped to read them.
The detection threshold drops as the cost of production collapses. Today, submitting an article represents a non-negligible time cost for a human researcher. This cost functions as a natural filter: one does not submit something one knows does not hold up. At fifteen dollars per submission, this brake disappears. The signal-to-noise ratio in the literature can degrade not because the produced articles are false, but because they are plausible without being interesting.
This distinction is crucial. A false article, a good reviewer detects it. A correct article, well-written, methodologically sound, but that contributes nothing, is much harder to reject — especially when reviewers themselves are under pressure from volume. The risk is not massive fraud. It is the dilution of the space within which important work must be found.
Precursor signals already exist. Several studies document a detectable increase in AI-assisted publications in machine learning conferences since 2023: systematic reviews of lexical patterns have identified common characteristics in hundreds of submissions, without the review process being able to effectively discriminate between these and entirely human articles. The phenomenon was modest in scale. It may not remain so.
Institutional irreversibility is the real subject here. Systems of scientific evaluation adapt slowly, through successive layers of norms, editorial practices, conventions. If peer review circuits adapt to processing volume without recalibrating their quality standard, this adaptation creates a path dependency difficult to undo. Institutions that have absorbed the flow by relaxing their criteria will not easily retrieve their original demands.
The Actors Building the Institutional Response
Facing this challenge, several concrete initiatives are underway, and their existence changes the picture.
Sakana AI itself has placed explicit limits on AI Scientist v2: the system cannot submit to journals without flagging its automated nature, and human authors remain formally responsible for submissions. This is not a technical constraint — it is a deliberate governance choice, inscribed in the system’s design.
Several major artificial intelligence conferences, including NeurIPS and ICML, updated their policies in 2024 and 2025 to require explicit declaration of the use of AI systems in writing and experiment design. These policies are imperfectly applied — no detection system functions at one hundred percent — but they create a normative framework that changes incentives.
The most structured initiative perhaps comes from the side of publication formats. Conferences are experimenting with submission formats that separate contributions: method on one side, interpretation and contextualization on the other. The idea is to make visible what an automated system can produce and what it cannot, by making both dimensions separately evaluable. This is a response of institutional design, not merely regulation.
Research in meta-science — the study of scientific production processes themselves — is also equipping itself with new tools. Teams are working on classifiers capable of identifying patterns characteristic of automated production, not to prohibit them, but to allow editors to treat them differently, with adapted evaluation protocols. The challenge is to maintain the discriminant capacity of the evaluation system in an environment of increased volume.
These responses are not yet equal to the challenge. But they exist and they are progressing, which is different from a situation where the institution would remain passive.
What This Changes for Researchers, Now
For a researcher working on machine learning, AI Scientist v2 is first of all a tool. Used as an assistant that proposes experiment variations, tests auxiliary hypotheses and produces a first draft of a methodology section, it frees time for what machines cannot yet do: choose problems worth posing, interpret results within a broader framework, make connections between fields.
This is a parallel with what calculation tools produced in physics in the 1960s and 1970s. The arrival of computers did not remove theoretical physics from physicists. It made certain computational tasks transparent, and freed cognitive capacity for conceptual work. The condition was that researchers know which questions to entrust to the machine and which to keep for themselves.
The same question arises today, and it is less trivial than it appears. In a discipline like machine learning, where a significant portion of academic production consists precisely in testing architectural variations and measuring their effects, the boundary between routine work and conceptual work is more porous than in other domains. This is where the real risk of displacement lies: not the replacement of the researcher, but the progressive reduction of the perimeter considered “interesting” to what systems can easily measure.
This question joins a broader debate about how digital tools reconfigure the nature of intellectual work. Telework showed that a tool can deliver routine work while impoverishing the exchanges that produce new ideas; the dynamic with scientific AI is structurally comparable. Maximizing the production of measurable articles does not necessarily maximize the production of useful knowledge.
The Tipping Point Has Not Yet Been Reached
AI Scientist v2 is a demonstration, not yet a massive deployment. Its domain of application remains narrow. Its cost, while low, assumes technical infrastructure and expertise to operate it. The fifteen dollars per article does not include the cost of engineers maintaining and directing the system.
But the trajectory is readable. The inference costs of large language models have fallen by roughly a factor of one hundred between 2022 and 2025, according to benchmarks published by Epoch AI. If this trend continues, even partially, an equivalent system could operate at a dollar or less per article by 2028. At that cost, the barriers to widespread adoption — including in contexts where the individual quality of submissions is little monitored — become very low.
This is why the window for institutional construction is short. The norms that will frame automated scientific production will be easier to establish before the volume is there than after. Conferences and journals that are now defining their AI transparency policies, that are experimenting with adapted submission formats, that are investing in detection and discrimination tools, are building an institutional asset whose value will increase as use of automated systems grows.
The question of who captures the gains from a new technology — and who absorbs the costs — arises here in its institutional version: if production savings benefit those who submit in mass, and if sorting costs are borne by a community of already overburdened volunteer reviewers, the distribution of gains is problematic. Part of the institutional response will need to address this imbalance explicitly.
Science has absorbed other profound mutations. The transition to open access publishing took two decades and remains incomplete. The reproducibility crisis, identified in the early 2010s, produced real changes in the practices of protocol registration and data sharing — slowly, imperfectly, but really. The arrival of systems like AI Scientist poses the following question: how quickly can scientific institutions learn, and is it fast enough this time?