Two AI systems have fundamentally challenged how science is done in 2026. Sakana AI's "The AI Scientist" became the first fully machine-generated research program published in Nature in March, after one of its papers passed peer review at one of the world's most prestigious AI conferences. Simultaneously, Google's AI co-scientist independently validated candidates for new liver fibrosis drugs, which were later confirmed in laboratory experiments.
Two systems, two approaches
"The AI Scientist" was developed by Sakana AI (Tokyo) together with researchers from the University of British Columbia, the Vector Institute, and the University of Oxford. The system independently generates hypotheses, searches scientific literature, designs and programs experiments, analyzes results, and writes complete research papers. In April 2025, it submitted three fully machine-written papers to a workshop at ICLR 2025, one of the world's most prestigious AI conferences. One paper was accepted, achieving a score of 6.33, better than 55 percent of human submissions.
Google's AI co-scientist takes a different approach. The multi-agent system built on Gemini 2.0 uses six specialized agents for generation, reflection, evaluation, evolution, similarity, and meta-evaluation. Its goal is not full autonomy but accelerating human research through iterative hypothesis generation and automated feedback.
Validated in the lab
Google's system identified new drug candidates for treating liver fibrosis through a collaboration with Stanford University. Stanford scientists then tested the suggestions in human liver organoids. The results, published in the journal Advanced Science: two of the AI-proposed substances significantly reduced fibrosis and promoted regeneration of liver parenchymal cells, all hits with p-values below 0.01. Similar results came from a search for new treatments for acute myeloid leukemia and from identifying a previously unknown mechanism for antibiotic resistance transfer.
Sakana AI's system currently works exclusively in machine learning as a research domain, because all its experiments are conducted computationally. It is not designed for field trials, laboratories, or physical apparatus. The system can also hallucinate literature citations and reaches its limits on methodologically complex questions. Sakana AI documents these limitations openly in the Nature publication.
What publication in Nature means
Nature's acceptance is a signal: the scientific community is beginning to treat AI-generated research as a serious part of the scientific process. Nature itself called in an accompanying commentary for institutions, research funders, and publishers to respond quickly to the new reality.
This raises structural questions that remain unanswered: Who is liable for flawed AI research? Who receives scientific credit? Can reviewers reliably detect that a paper is machine-generated? A paper scoring better than more than half of all human submissions suggests the dividing line is already blurrier than many assume.
The liver experiment also shows how AI co-scientist systems can work in practice: not as replacements for human scientists, but as tools that generate hypotheses which human teams then test and evaluate. The difference is substantial. Instead of months of literature work, a system delivers candidates in hours that can be verified in the lab.
What comes next
OpenAI has announced plans to develop an AI research intern capable of supervised research by September 2026, and a fully autonomous AI researcher by March 2028. Sakana AI is working on a follow-up version of "The AI Scientist" that can conduct research beyond machine learning into other fields. Nature has already called in an editorial for conferences and journals to develop specific disclosure requirements for AI-generated research contributions. No binding standard exists yet.