Google Publishes AI System in Nature, Controls Who Gets Early Access
Google Controls Which Science Gets Published First
Google published a scientific AI system in Nature four days ago and immediately decided which researchers could use it first. ERA — Empirical Research Assistance — can devise and implement novel scientific methods faster than most labs. But the more consequential question was already settled before the paper appeared: which scientists get early access, which findings reach publication first, and what questions get investigated at all. On all three, the answer is Google.
ERA, described in a Nature paper published May 19, 2026, combines a large language model with a search algorithm called Tree Search to explore thousands of possible approaches to a scientific coding problem in hours or days, rather than the months a human researcher would typically spend. In single-cell data analysis — a technique biologists use to understand how individual cells differ from one another — ERA discovered 40 methods that performed well on a public leaderboard. In epidemiology, it generated 14 models that outperformed the Centers for Disease Control and Prevention's forecasting ensemble for predicting COVID-19 hospitalizations, according to Google Research Blog. It also ranks at or near the top of public CDC leaderboards for flu and RSV hospital admissions forecasting.
The system is not waiting for academic evaluation. ERA is available through a trusted tester program in Google Labs via Gemini for Science — Google's experimental tool for computational discovery — according to the Google Research Blog. Eight manuscripts already apply ERA to specific scientific problems, including epidemiological forecasting, California seasonal runoff prediction, and atmospheric CO2 mapping. Michael Brenner, a co-lead on the project and both a Harvard professor and a Google research scientist, put it plainly in a Harvard SEAS news release: the system automates coding for scientific research. The work involved Harvard PhD students Qian-Ze Zhu, Ryan Krueger, and Sarah Martinson as Google student researchers.
The trusted-tester structure is not a neutral distribution channel. It means Google controls which researchers get access, which scientific questions get investigated first, and — by selecting among the eight active manuscripts — which findings reach publication before others. A company that both produces a scientific instrument and selects who gets to use it early is making editorial decisions about the order of discovery. That is a different kind of gatekeeper than a corporation selling a tool that researchers choose to adopt on their own timeline.
The epistemological stakes follow from that control. Every scientific code generation embeds assumptions: what optimization objective to prioritize, which statistical frameworks to prefer, which kinds of hypotheses merit exploration. Those are methodological choices that human researchers make based on disciplinary training, intuition, and implicit scientific judgment. ERA makes those choices automatically, optimizing for what the leaderboard measures — and the leaderboard reflects what researchers chose to measure. The loop is closed, but it is not independent. A system that searches the space of possible approaches fastest does not ask whether the space itself is the right one to search.
The Nature paper is authored by Google employees. The performance claims — 40 methods, 14 epidemiology models — are validated against public leaderboards, which is the best available standard, but the public leaderboards themselves were built by the scientific community Google is now selling into. The GitHub repository for ERA has 168 stars and 27 forks, with the last commit May 20, 2026 — modest organic traction for a system that Google is simultaneously publishing in Nature and distributing through its own product channel. Independent replication of the specific claims has not been reported.
What makes the "devised and implemented" language in the abstract worth sitting with: it is not the language of a tool. Tools assist. Authors create. The paper's choice of words suggests ERA is doing something closer to the latter — and if that claim holds up, the more important story is not the benchmark scores. It is the question of who is responsible for the hypotheses a system generates, who gets to set the agenda for what science gets attempted next, and what assumptions get embedded in the code that produces the record.
On all three questions, the answer right now is Google.
The broader pattern fits something the scientific community has been tracking for years: as AI systems become capable of designing experiments, selecting datasets, and writing the code that produces scientific results, the question of who controls the methodology is increasingly a question about who controls the science. ERA is not the first system to automate a step in the research pipeline. It is one of the first to package that automation as a product — and to distribute it through a trusted-tester program that Google operates unilaterally. The benchmarks are the easy part to verify. The harder thing to track is what questions researchers stop asking because ERA answered them first.