A system of artificial intelligence developed by Sakana AI in collaboration with Canadian and British researchers has demonstrated its ability to conduct scientific research from ideation to article writing. Published in Nature, this system successfully submitted an article to an academic conference, marking a significant advancement in research automation.
A system capable of autonomously conducting scientific research, from idea generation to drafting complete articles, represents a milestone in the evolving relationship between artificial intelligence and the scientific method. Developed by Sakana AI in collaboration with researchers from the University of British Columbia, the Vector Institute, and the University of Oxford, this system named AI Scientist functions as a complete virtual researcher.
A scientific process automation: The mechanism relies on foundational models that orchestrate each phase of scientific work. It generates research ideas, explores academic literature to verify the originality of proposals, writes and corrects code for experiments, analyzes results, produces data visualizations, drafts manuscripts in LaTeX, and evaluates the quality of its own production. “This article marks the dawn of a new chapter in human history, where scientific progress is radically accelerated by AI scientists capable of autonomous action,” emphasizes Jeff Clune, professor of computer science at UBC and lead author of the publication.
The team has also developed an automated evaluator capable of predicting acceptance decisions at conferences with performance comparable to human evaluators. This component has led to the establishment of what researchers describe as a scalability law: the quality of produced articles improves proportionally to the capabilities of underlying foundational models and allocated computational power.
Testing under real conditions: To evaluate the system’s performance against academic standards, researchers submitted three articles entirely generated by artificial intelligence to a workshop at the International Conference on Learning Representations in 2025. One of these articles, focusing on neural network regularization, received an average score of 6.33 out of 10 from human evaluators. This performance ranked above approximately 55% of all submissions and exceeded the workshop’s acceptance threshold.
In accordance with an agreement with the conference organizers, Sakana AI withdrew the article before publication, citing the absence of established standards for AI-generated manuscripts.
Current capabilities and limitations: Researchers acknowledge several shortcomings in the current system: – It sometimes produces underdeveloped ideas – It generates inaccurate citations – It is currently limited to research in computer science
Sakana AI itself indicated that none of its three submissions to ICLR met internal standards for publication in the conference’s main session. The accepted article only passed through a workshop session with an acceptance rate of 60 to 70 percent.
Despite these limitations, the potential implications of this technology have attracted the attention of the scientific community. “AI Scientist paves the way for recursive improvement in which the AI system not only discovers new scientific knowledge but uses these discoveries to become better at realizing further discoveries,” explains Shengran Hu, a UBC doctoral student and study co-author.
An editorial in Nature published concurrently with the scientific article emphasizes that the system “raises unanswered questions about how research should be conducted and governed as AI-driven automation accelerates.”
The emergence of systems capable of automating the entire scientific process poses fundamental challenges for the future of research. While the technology accelerates certain phases of scientific work, it also raises questions about the role of human intuition, creativity, and responsibility in knowledge production.
(Note: Paper citation provided at the end of the article for reference)
