SScience has always relied on the inquisitive human mind to formulate hypotheses, design experiments, analyze results, and present cases to colleagues. Over the centuries, we’ve built better tools like electron microscopes, particle accelerators, and supercomputers, but the core loop of scientific discovery remains stubbornly human. Now, for the first time, the loop has begun with a new feeling.
Until now, scientists have often relied on artificial intelligence to help them solve narrow, predefined tasks, such as folding proteins, said Jeff Clune, a computer science professor at the University of British Columbia. “We’re saying AI can be a scientist,” he says.
Recently nature In this study, Clune et al. introduced AI Scientist. The AI system wrote the paper without human intervention and passed peer review at the 2025 International Conference on Learning and Representations (ICLR) workshop, the premier venue in the field of machine learning. Clune and other experts say the paper was mediocre. But its existence marks a tipping point that the scientific community is just beginning to grapple with. AI has quickly moved from assisting scientists to becoming scientists.
About supporting science journalism
If you enjoyed this article, please consider supporting our award-winning journalism. Currently subscribing. By subscribing, you help ensure future generations of influential stories about the discoveries and ideas that shape the world today.
AI Scientist consists of multiple modules. When given a general topic prompt by a researcher, we examine the available literature and generate a hypothesis. “We’re just giving them a general direction, like, ‘Come up with something interesting to study about how AI learns,'” Clune explains. The system then evaluates and refines those ideas, eliminating those that are not novel. From there, further modules plan and run experiments, analyze and plot data, and finally write the paper. Clune said the company also uses its own internal peer review process to find flaws in papers. (The system relies on existing underlying models, such as Anthropic’s Claude Sonnet and OpenAI’s GPT-4o. The team’s contribution is a pipeline that tunes these models.)
To see if The AI Scientist’s work meets human standards, the team submitted three papers generated by the AI Scientist to the “I Can’t Believe It’s Not Better (ICBINB)” workshop at ICLR in 2025. One person was accepted. (Conference organizers allowed AI-generated papers to be submitted, and all papers by AI scientists were withdrawn from the conference after a review process.)
The team behind AI Scientist admits that the hurdles for this workshop were lower than for major conference publications. “How can a mediocre graduate student get one in three papers accepted at a university that accepts 70 percent of papers? Absolutely!” says Jody Schneider, an associate professor of information science at the University of Wisconsin-Madison who was not involved in Clune’s research.
The AI paper is “okay, but not great,” Clune said. To him, some of the AI ideas seemed really creative, but the system was struggling to implement. “The logic, writing, and ideas throughout the paper did not all fit together beautifully,” he points out. Additionally, there were issues such as hallucinatory references, duplicate figures, and lack of methodological rigor.
Overall, Kroon et al.’s new study has received lukewarm reviews. “This approach is agentic and has no real novelty,” says Maria Liakata, a professor of natural language processing at Queen Mary University of London, who was not involved in the study.
However, there was one indicator where AI scientists significantly outperformed human researchers. Clune wrote an officially passable paper on machine learning in less than 15 hours, at an estimated cost of about $140. Compare this to the abilities of a graduate student, Schneider said. Graduate students may take a semester to write their first accepted workshop paper.
As costs fall and output speeds increase, AI-authored papers pose pressing challenges to the scientific community. “Articles written by AI will probably make the situation even worse,” warns Yanan Sui, an associate professor at China’s Tsinghua University and chair of the ICLR 2026 senior workshop.
To protect themselves from this flood, top venues are starting to put restrictions in place. “The main conference has strict rules and does not allow the submission of papers written purely using AI,” Sui said. The current compromise is forced transparency. Authors using AI must clearly state how the AI was used. However, Sui acknowledges that journals and conferences typically lack the tools to reliably detect AI-generated submissions.
Meanwhile, tools to autonomously create these posts are already becoming popular. Intology claimed that its AI Zochi passed peer review of the main proceedings of the Association for Computational Linguistics’ 63rd Annual Meeting (although human researchers were involved in areas such as validating results before submission and communicating with reviewers). Another group called the Autoscience Institute says a paper produced by its AI system was accepted at the ICLR workshop before AI Scientist.
“You can’t take away the power of AI to produce scientific papers,” says Aaron Shine, a data scientist at the University of Chicago and one of the organizers of the ICBINB workshop. “This technology will only continue to improve. I don’t think there’s anything we can do about it.”
But what happens when one day the AI-generated paper stops being mediocre?
Clune sees the transition unfolding in two stages. “In the very short term, there will be a lot of waste and trash, and the peer review system will have to deal with that,” he says. But ultimately, he argues, AI systems will be far better at science than human researchers. “I predict that AI scientists will indeed mark the dawn of a new era of rapid scientific progress,” Clune argues, imagining humans taking on the role of curators witnessing AI accomplish scientific wonders.
But Liakata believes we humans still have work to do. “I believe the future is not fully autonomous scientific discovery, but advanced human-agent interaction where humans can probe and contribute to the process,” she says.
