Science has always started with a human asking a question. That assumption just became considerably less reliable.
A study published in Nature has documented the first AI system to autonomously complete the full research cycle - generating ideas, running experiments, analyzing data, writing a manuscript, and passing peer review at a major machine-learning conference workshop. The system, called The AI Scientist, was developed by Sakana AI, a Tokyo-based research company.
The paper was submitted to the I Can't Believe It's Not Better workshop at ICLR 2025 and passed. It was produced in 15 hours at a cost of approximately $140.
What The AI Scientist Actually Does
The system is not a single model but a pipeline of coordinating AI modules built on top of existing foundation models including Anthropic's Claude Sonnet and OpenAI's GPT-4o. Researchers give it a general topic direction - something like "come up with something interesting to study on how AI learns" - and the system takes over from there.
It surveys available literature, generates and filters hypotheses for novelty, plans and executes experiments, analyzes and visualizes data, writes the full manuscript, and then conducts its own internal peer review to identify weaknesses before submission. The human contribution is the initial prompt and the pipeline architecture. The research content is entirely machine-generated.
Why the Peer Review Passage Matters
Passing peer review is meaningful precisely because it is the mechanism science uses to distinguish credible work from noise. A system that can produce formally passable research papers at $140 per paper and 15-hour turnaround represents something qualitatively new - not just a productivity tool, but a system that can mimic the outputs of scientific expertise at scale.
Jeff Clune, a computer science professor at the University of British Columbia who worked on the project, framed the near-term reality honestly: in the very short term, there will be a lot of slop and garbage, and peer review systems will have to deal with it. His longer-term prediction is starker - that AI systems will eventually be far better at science than human researchers.
The Threat to Scientific Infrastructure
The concern that most experts are raising is not about this paper specifically. It is about volume. AI can generate research infinitely faster than humans can read it. The peer review system is already under severe strain - too many submissions, too few willing reviewers, declining satisfaction with review quality. An influx of AI-generated papers threatens to overwhelm a system that was already breaking.
Top-tier venues like ICLR have responded by setting strict rules prohibiting purely AI-written submissions at main conferences. But conference workshops, where this paper passed, operate under different standards. And the rules are easier to state than to enforce - AI detection tools remain unreliable, and the same AI writing quality that can mask weak ideas also makes automated detection harder.
Sakana AI has called for clear disclosure standards as a minimum safeguard. Whether the scientific community can agree on and enforce those standards fast enough to keep pace with the technology is an open question with no obvious answer.
