Sakana AI Publishes in Nature: An AI That Conducts Scientific Research End-to-End

Sakana AI, in collaboration with the University of British Columbia, the Vector Institute, and the University of Oxford, published an open-access paper in Nature on March 26, 2026, detailing its "AI Scientist" system — an artificial intelligence capable of automating the entire machine learning research lifecycle.

Key Highlights

• Published in Nature: the paper comprehensively details the system's architecture, scaling results, and the challenges of AI-generated science
• First fully AI-generated paper to pass peer review: a manuscript produced by AI Scientist-v2 scored an average of 6.33 at the ICLR 2025 ICBINB workshop, surpassing the average human acceptance threshold
• Scaling law discovered: as the underlying foundation models improve, the quality of generated papers increases correspondingly
• Fully open source: code for both v1 and v2 is available on GitHub

How AI Scientist Works

The system autonomously executes every step of the research process:

1. Ideation: generates novel research hypotheses
2. Literature review: searches and reads relevant papers
3. Experimentation: designs, programs, and runs experiments via parallelized tree search
4. Writing: produces complete papers in LaTeX with vision-based figure feedback
5. Evaluation: an Automated Reviewer achieves 69% balanced accuracy, exceeding inter-human agreement levels measured at NeurIPS 2021

Current Limitations

Despite these advances, the system still has notable shortcomings. It occasionally produces naive or underdeveloped ideas, struggles with methodological rigor and complex coding, and remains susceptible to hallucinations, inaccurate citations, and duplicated figures. As one commenter on X noted: "AI crushes the railroad efficiency... but real breakthroughs need the open-road collisions only humans and AI together can make."

Why It Matters

This publication marks a significant milestone for automated research. The discovery of a scaling law — where paper quality improves directly with foundation model capability — suggests future iterations will be substantially more capable.

The researchers took notable ethical precautions: watermarking AI-generated papers for transparency, obtaining IRB approval for experiments, and voluntarily withdrawing the accepted paper from the conference to avoid overwhelming the peer review system.

What's Next

Sakana AI envisions accelerated scientific discovery across multiple domains — from medicine to environmental protection. However, the company emphasizes the need for community norms around how AI-generated research is treated, to prevent the scientific publishing system from being flooded with automated submissions.

Source: Sakana AI