Autonomous AI Scientist: Full Research Cycle from Idea to Publication
The AI Scientist system from Sakana AI has for the first time completed the entire scientific research cycle without human intervention: from hypothesis generation to writing the manuscript and self-evaluation. Nature journal published a report on this, highlighting the value of the process despite the negative experimental result. The AI chose the topic of compositional regularization to improve neural network training, tested the method, and concluded there was no effect. This approach allows publishing negative results, saving community resources.
The AI Scientist uses an agent-based architecture based on GPT-4, Claude, and Llama 3. The cost of one full article is about $15, making the system scalable for mass-testing ideas.
System Modes
The system supports two modes:
- Focused: the developer sets a code template and topic, and the AI generates ideas within those constraints.
- Free: autonomous agent-based search through scientific literature to select directions.
In both cases, the process includes:
- Hypothesis generation.
- Writing and running experiment code.
- Data analysis, plotting graphs.
- Forming a manuscript with abstract, introduction, methods, and conclusions.
- Self-review.
Out of three generated articles, one received an average score of 6.33 at the ICLR 2025 workshop, but was withdrawn before publication by agreement with the organizers.
Challenges for the Scientific Community
Nature warns of risks: cheap article generation could overload peer review and increase the share of low-quality content. In 2025, cases of AI hallucinations in references were already recorded. Sakana AI, OpenAI, Google DeepMind, and Anthropic are competing in developing similar systems, but autonomous AI articles are not yet accepted in main conference tracks or journals.
Professor Jeff Clune from UBC noted the potential to accelerate progress but emphasizes the need for ethical standards.
The distinction between AI as a tool and AI as an independent author remains blurred and requires community discussion.
Key Points
- The AI Scientist is the first system published in Nature for an autonomous scientific cycle.
- The negative result on compositional regularization was published as a method demonstration.
- Research cost — $15, using GPT-4, Claude, Llama 3.
- One article passed review at the ICLR 2025 workshop (score 6.33), but was withdrawn.
- Risks: overload of reviewing and growth of AI-generated noise in the literature.
— Editorial Team
No comments yet.