On 25 March 2026, Nature published “Towards end-to-end automation of AI research,” describing The AI Scientist, an agentic system that can propose ideas, search the literature, write code, run machine-learning experiments, analyse results, draft a complete paper and even perform automated peer review. Its most arresting result is not that the science was revolutionary, but that one fully AI-generated manuscript cleared the first round of blind review at the ICLR 2025 I Can’t Believe It’s Not Better workshop. That paper earned scores of 6, 7 and 6, ranked in the top 45% of submissions sent for review, and reported a negative result. Still, the triumph should not be overstated: only one of three AI-generated submissions was accepted, and the workshop’s acceptance rate was 70%, far above the 32% reported for the ICLR 2025 main conference. (nature.com)
What makes the study important is its architecture. The system was tested in two modes: a focused mode built on human-provided code templates, and a more open-ended mode that writes initial code itself and explores alternatives through agentic tree search. Because the work is confined to machine learning, the entire research loop can unfold inside a computer, making full automation unusually plausible. To judge output at scale, the authors also built an “Automated Reviewer,” which followed NeurIPS-style reviewing rules and achieved balanced accuracy comparable to human reviewers on public ICLR decisions: 69% on earlier data and 66% on a post-cutoff 2025 set. The paper further reports a clear trend: stronger underlying models and more test-time compute tend to yield better AI-written papers. (nature.com)
So, can AI write papers and pass peer review? In a narrow sense, yes. In the broader sense that matters to science, not yet. The authors explicitly list recurring weaknesses: naïve ideas, weak methodological rigour, implementation errors, duplicated figures and hallucinated citations. They also warn that such systems could flood peer review, inflate credentials, misappropriate ideas and add noise to the literature. Nature’s editorial makes the same larger point: institutions, funders and publishers will have to rethink the rules of science itself. Perhaps that is the real shock of The AI Scientist. The question is no longer whether AI can imitate parts of research, but whether academia is prepared for a world in which imitation becomes institutionally consequential. (nature.com)










