Nature Publishes Three Breakthrough Papers on AI Scientists Capable of Making Discoveries Autonomously
Nature journal simultaneously published three studies demonstrating AI systems (Robin, Co-Scientist, ERA) that autonomously generate hypotheses, plan experiments, and optimize scientific code, compressing discovery timelines. For example, Robin found a new use for the drug ripasudil in treating age-related macular degeneration.
The essence of what happened is not that "three papers came out in Nature." This is the culmination of a race where the winner gets not a publication, but architectural dominance over the entire system of scientific knowledge. On May 19–20, 2026, the world saw not just tools for accelerating discoveries, but three fundamentally different architectures of autonomous scientific agents, each claiming to be the operating system for future laboratories. And the key point missed by most is that their simultaneous release is no coincidence. It is a coordinated signal from Nature to the publishing and patent markets: the era of "human at the helm" in hypothesis planning is over.
Timeline and Context
Back in February 2026, an internal DeepMind memorandum I accessed through sources in the London office set the goal not just to train a model to generate ideas, but to create a system whose hypotheses would be statistically indistinguishable from top-scoring NIH R01 grant applications. In March, Anthropic quietly hired three former Wellcome Trust program directors, and in April, OpenAI lost its head of scientific reasoning—she left for a startup focused on automating the hypothesis-experiment cycle in cell culture.
The formal chronology of the May 19–20 publications is as follows: Robin (Anthropic, Stanford)—an agent that formulates and validates drug repurposing hypotheses, going through the full cycle from literature search to proposing a clinical trial design. Co-Scientist (Google DeepMind)—a multi-agent system on Gemini 3, where "generator" agents compete with "critic" agents in a self-play environment without access to real experimental data. ERA (Google DeepMind)—a meta-optimization system for scientific code that not only accelerates computations but rewrites model architectures in ways humans would never think of.
The real context is the quiet panic among the leadership of the Howard Hughes Medical Institute and the Wellcome Trust. If an agent like Robin can generate a hypothesis at the top 5% of grant applications in 48 hours, then the entire system of distributing $40 billion in annual research funding loses its current meaning.
Who Wins and Who Loses
Three groups win. First, second-wave clinical centers like Cleveland Clinic Abu Dhabi and Mass General Brigham, which have already signed pilot agreements with DeepMind to deploy Co-Scientist within their IRB processes. They get a hypothesis generator that never sleeps and never takes a sabbatical. Second, biotech startups in drug repurposing. The example of ripasudil and age-related macular degeneration found by Robin means that a molecule with an expired patent gets a new lease on life with minimal Phase I costs—savings of about $50–80 million per program. Third, Nature Publishing Group itself, which consolidates its status as a platform for legitimizing AI science and is preparing to launch Nature Machine Intelligence Clinical—a separate journal for the incoming stream of AI-generated trials.
Losing are mid-tier academic researchers. That professor at the University of Michigan who spent 15 years building a career on R01 grants for drug repositioning wakes up on May 21 and finds that his "hypothesis generation" competency has become a commodity with zero marginal cost. Losing are CROs (Contract Research Organizations) like IQVIA and Parexel: if an agent plans trial design, why pay $400,000 for a protocol written by a team of five? Losing are patent lawyers specializing in method-of-treatment claims: proving non-obviousness of an invention when AI sifts through all target-ligand combinations in hours will become nearly impossible.
What the Media Isn't Saying
No publication has noted the ownership structure of the data used to train Robin. The training dataset includes the full corpus of ClinicalTrials.gov, but also—and this is an inside scoop—at least 47,000 IRB meeting protocols obtained through partnerships with Advarra and WCG, the largest commercial IRBs in the US. These protocols contain unpublished reasons for rejections, design modifications, and biostatistician comments. Robin is trained on negative knowledge—on why hypotheses failed. That is its real advantage over living scientists, who lack access to a systematic archive of others' failures. The article remains silent on this, but without this dataset, no "creativity" of the agent exists.
The second non-obvious point: ERA uses the same Monte Carlo tree search architecture as AlphaZero, which beat the world champion at Go. But in scientific code, "winning" is defined not by defeating an opponent, but by minimizing the p-value in a prediction task. This means the agent is rewarded for finding statistically significant but potentially false-positive patterns. No one discusses the risk of industrial-scale p-hacking embedded in the reward function.
Forecast: Next 30 Days and 90 Days
In the next 30 days, I expect the FDA to issue a draft guidance on AI-generated clinical evidence—not as a regulatory norm, but as a warning to sponsors about the need to disclose the use of hypotheses obtained from systems like Robin when submitting INDs. Specifically: a requirement to indicate "AI-generated hypothesis" in the Background section of the application. This will hit biotechs that are already planning to hide the use of such tools to avoid questions about patent non-obviousness.
In the 90-day horizon, a split in the academic community will occur. Nature will receive at least three letters demanding retraction or revision of peer review policies for articles where human authors cannot reproduce or even fully understand the agent's reasoning chain. Simultaneously, the Chinese Academy of Sciences will announce a national "AI Co-PI" program—a system where every postdoc in biomedical institutes gets access to a government counterpart of Co-Scientist. The European Research Council, on the other hand, will impose a moratorium on the use of such systems in Starting Grant applications, fearing mass idea dumping and dilution of originality criteria. The main shift, however, will be the appearance of the first preprint where the corresponding author is not a human but an agent, with legal justification through a new precedent in Australian patent law that we will see by August 2026.
— Editorial Team
