By Connie · This article contains affiliate links. We may earn a commission at no extra cost to you if you sign up through our links.
AI Scientist-v2 Published in Nature: The First Autonomous AI Research Paper to Survive Peer Review
Sakana AI's AI Scientist-v2 is the first AI system to autonomously write a full research paper and get it published in Nature. The system costs $6–$15 per paper and requires just 3.5 hours of human oversight. Academic conferences are now scrambling to respond.
A machine just published a paper in Nature. Not with human ghostwriters or light AI editing — but as a fully autonomous system that read the literature, designed experiments, ran the code, analyzed results, and wrote the manuscript from scratch.
Sakana AI's AI Scientist-v2 crossed the threshold that the research community long treated as a distant horizon. On March 25, 2026, Nature published the paper describing the system — and the world noticed.
How AI Scientist-v2 Works
The system is an agentic loop. It starts with a research question, browses existing literature, proposes testable hypotheses, writes and executes experimental code, interprets the output, and produces a formatted paper complete with figures and citations.
The key upgrade in v2 is an agentic tree search. Rather than committing to a single experimental path, the system branches across multiple hypotheses simultaneously, scores them by promisingness, and prunes dead ends. This dramatically increases the rate of producing publishable-quality ideas.
Built by Sakana AI in collaboration with the University of British Columbia, the Vector Institute, and the University of Oxford, the system represents roughly 18 months of research effort — all of which now fits in a pipeline that can run overnight.
The Path to Nature
The story actually starts in 2025. AI Scientist-v2 submitted three papers to the ICLR 2025 workshop "I Can't Believe It's Not Better." One paper — on compositional regularization in neural networks — was accepted with an average score of 6.33. Sakana AI withdrew it before publication, partly to avoid controversy before the work was fully documented.
The Nature publication consolidates everything: the system design, the methodology, the results, and an honest accounting of limitations. Nature's peer reviewers accepted it — but forced the team to significantly walk back earlier claims of full automation. The final paper acknowledges that humans helped filter promising outputs and expand the discussion of failure modes.
Performance Numbers
| Metric | Value |
|---|---|
| Cost per paper | $6–$15 compute cost |
| Human involvement | ~3.5 hours (oversight, not writing) |
| Experiment failure rate | 42% (due to coding errors) |
| Citation hallucination | Occasional (documented in audit) |
| Scope | Computational ML only — no wet lab or field work |
| First peer-reviewed acceptance | ICLR 2025 workshop (later withdrawn) |
| Nature publication | March 25, 2026 |
Why This Matters — and Why It's Complicated
The obvious implication: if a machine can produce publishable research for $15, the economics of academic science are about to shift. Labs that once needed teams of PhD students to run literature surveys and preliminary experiments can now delegate that work to an autonomous agent.
The less obvious implication: the scientific literature is about to face a flood. Researcher Jevin West described the risk as an overwhelming volume of low-cost, AI-generated submissions that bury the genuinely novel work humans produce. ICLR 2026 is already responding — banning pure AI-generated papers from its main proceedings and requiring authors to disclose all AI assistance.
The independent audit by Joeran Beel's team is the most grounding data point. Forty-two percent of proposed experiments failed due to coding errors. The system hallucinated citations. It sometimes duplicated figures. These are not trivial issues — they are the difference between a tool that accelerates research and one that pollutes it.
Sakana AI's honest response to the audit is a sign that the team understands the stakes. The Nature publication includes the limitations. The system is real, but it is not a replacement for scientific judgment — it is an accelerant for people who already have it.
What Comes Next
Sakana AI has not announced a commercial release of AI Scientist-v2. The current system is limited to machine learning experiments — domains where the "wet lab" is a GPU cluster. Extending it to biology, chemistry, or social science would require entirely different infrastructure and raises much harder questions about experimental safety.
The academic community is already moving. Major conferences are drafting emergency disclosure policies. Some journals are considering AI-specific submission categories. The peer review process itself — already under strain from volume — is being reconsidered from first principles.
The bigger question is what happens when every university lab has access to a system like this. The bottleneck in science has always been human time — time to read, design, execute, and write. AI Scientist-v2 removes most of that bottleneck. What remains is the hardest part: knowing which questions are worth asking.
How Multi-Model AI Platforms Help Researchers Today
While AI Scientist-v2 is a specialized research system, everyday researchers and knowledge workers are already using multi-model AI platforms like Happycapy to compress their research cycles. The ability to query Claude for deep reasoning, GPT-4 for broad knowledge synthesis, and Gemini for multimodal analysis — all within a single session — reflects the same underlying shift: AI as a research accelerant, not a replacement.
See also: How to Use AI for Research in 2026 and GPT-5.4 vs Claude Opus 4.6 vs Gemini 3.1 Pro.
Get the best AI tools tips — weekly
Honest reviews, tutorials, and Happycapy tips. No spam.