A new generation of artificial intelligence is moving beyond assisting researchers and is now capable of independently completing full scientific studies from start to finish, according to a breakthrough paper published in Nature. The system, described by an international research team, can generate original ideas, conduct experiments, analyze results, write academic papers, and even evaluate its own work with minimal human involvement.
The project was developed through a collaboration between scientists at University of British Columbia Computer Science, Sakana AI, the Vector Institute, and the University of Oxford. Researchers say the system represents the first demonstration of an AI capable of autonomously completing the entire scientific workflow, marking a significant milestone in how artificial intelligence could reshape research and innovation.
Unlike earlier AI tools that supported specific steps such as analyzing medical scans or predicting protein structures, the new system can independently carry out the full research cycle. It begins by proposing novel research questions, checking existing literature to confirm originality, writing and debugging experimental code, analyzing results, generating visualizations, drafting manuscripts, and performing its own evaluation process. The researchers built the system using foundational language models similar to widely known tools such as ChatGPT.
To test whether the system could meet academic standards, the team submitted a fully AI-generated research paper to a workshop connected with the International Conference on Learning Representations. The submission successfully passed peer review, demonstrating that the AI’s output could meet expectations typically applied to human researchers. In addition, the team created an automated reviewer capable of predicting acceptance decisions for conference submissions with accuracy comparable to human evaluators, further strengthening confidence in the system’s research capabilities.
Researchers say one of the most promising aspects of the technology is its potential for recursive self-improvement. The system could eventually refine its own methods using insights from its discoveries, allowing it to become progressively more capable over time. That possibility opens the door to a new model of scientific progress in which AI systems continuously expand their own knowledge base and research performance without direct human intervention.
Despite the breakthrough, the team also identified several limitations. The AI scientist occasionally produced incomplete ideas and inaccurate citations, and its current capabilities are largely restricted to computer science research tasks. However, researchers believe future versions could extend into other disciplines such as biology, physics, and engineering as models become more advanced and computing resources increase.
Looking ahead, scientists involved in the project suggest that entire collaborative networks of AI researchers could eventually emerge, functioning similarly to human academic communities. In such a scenario, discoveries generated by one system could inform the next round of experiments by another, creating a continuous cycle of innovation that could dramatically accelerate the pace of scientific advancement worldwide.
