AI wrote a scientific paper that passed peer review

SScience has always relied on the curious human mind to formulate a hypothesis, design an experiment, analyze the results, and present the case to peers. Over the centuries we have built better tools such as electron microscopes, particle accelerators and supercomputers, but the heart of scientific discovery has remained stubbornly human. Today, for the first time, this loop begins with a new type of mind.

Until now, scientists have often benefited from the help of artificial intelligence to solve a predefined and precise task, such as protein folding, says Jeff Clune, professor of computer science at the University of British Columbia. “We say that AI must become the scientist,” he says.

In a recent Nature In this study, Clune and his colleagues unveiled AI Scientist, an AI system that wrote a paper without human intervention and passed peer review for a workshop at the 2025 International Conference on Learning Representations (ICLR), a leading venue in the field of machine learning. The document was poor, according to Clune and other experts. But its existence marks a turning point that the scientific community is only beginning to come to grips with: AI has quickly moved from being an aid to scientists to attempting to be one.

On supporting science journalism

If you enjoy this article, please consider supporting our award-winning journalism by subscription. By purchasing a subscription, you are helping to ensure the future of impactful stories about the discoveries and ideas shaping our world today.

The AI ​​Scientist includes several modules. After researchers give him a general topic, he reviews the available literature and generates hypotheses. “We just give it a general direction like, ‘Come up with something interesting to study about how AI learns,’” Clune says. The system then evaluates and refines these ideas, filtering out those that are not new. From there, other modules plan and run experiments, analyze and plot the data, and finally write the paper. It even conducts its own internal peer review process to find flaws in its papers, Clune says. (The system builds on existing foundation models such as Anthropic’s Claude Sonnet or OpenAI’s GPT-4o; the team’s contribution is the pipeline that orchestrates these models).

To see if The AI ​​Scientist’s results could meet human standards, the team submitted three papers generated by it to the I Can’t Believe It’s Not Better (ICBINB) workshop at ICLR 2025. One was accepted. (Conference organizers gave permission for AI-generated papers to be submitted, and all papers by AI scientists were removed from the conference after the review process.)

The team behind AI Scientist admits that the bar for this workshop was lower than that of a keynote conference publication. “Would a mediocre graduate student get one in three papers accepted at an institution that accepts 70 percent of papers? Of course!” says Jodi Schneider, associate professor of information sciences at the University of Wisconsin-Madison, who was not involved in Clune’s study.

The AI ​​papers “are OK but not great,” Clune says. To him, some of the AI’s ideas seemed genuinely creative, but the system struggled to execute. “The logic, writing and thinking throughout the document did not fit together perfectly,” he notes. Other problems involved wild references, duplicate figures and a lack of methodological rigor.

Overall, Clune and colleagues’ new study received a lukewarm reception. “The approach is agentic and without any real novelty,” says Maria Liakata, professor of natural language processing at Queen Mary University of London, who was not involved in the work.

There is, however, one metric in which the AI ​​scientist vastly outperformed human researchers: He produced a formally passable paper on machine learning in 15 hours, at a cost Clune estimated at around $140. Compare that with the ability of a graduate student, who might take a full semester to write their first accepted workshop paper, according to Schneider.

As costs fall and production speeds increase, AI-authored papers present the scientific community with an immediate challenge. “AI-authored papers will likely make things worse,” warns Yanan Sui, associate professor at Tsinghua University in China and lead chair of the ICLR 2026 workshop.

To guard against this flooding, prominent sites have started setting limits. “There are strict rules for the main conference that do not allow purely AI-written papers to be submitted,” says Sui. The trade-off, for now, is forced transparency: authors using AI must clearly state how it was used. Sui admits, however, that journals and conferences generally lack the tools to reliably detect AI-generated contributions.

In the meantime, tools for writing these contributions independently have already begun to proliferate. Intology claimed that its Zochi AI passed peer review for the main proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (although human researchers were involved in areas such as verifying results before submission and communicating with peer reviewers). Another group called Autoscience Institute said its AI system created papers that were accepted at ICLR workshops before the AI ​​Scientist.

“We’re not going to be able to take away the power to generate scientific papers about AI,” says Aaron Schein, a data scientist at the University of Chicago and one of the organizers of the ICBINB workshop. “This technology is only going to get better. I don’t think there’s anything that can be done about it.”

But what if one day AI-generated logs stopped being mediocre?

Clune sees the transition taking place in two phases. “In the very short term, you’re going to have a lot of trash and garbage, and peer review systems are going to have to deal with that,” he says. But eventually, he says, AI systems will be much better at science than human researchers. “I predict that AI Scientist actually marks the dawn of a new era of rapid scientific progress,” says Clune, imagining humans reduced to curators watching AI perform scientific wonders.

Liakata, however, believes that we humans still have something to do. “I believe the future is not one of completely autonomous scientific discovery, but rather advanced human-agent interaction where humans can examine and contribute to the process,” she says.