This is where ERA differs. ERA does not stop after writing a program once. A large language model generates candidate code, the programs are executed using actual data, and each candidate is scored according to predetermined evaluation metrics. The system then revises the code around the higher-scoring candidates, creates new combinations, and explores directions that improve performance. While the language model proposes ideas and code, tree search determines which candidates deserve deeper exploration.

This process resembles the way researchers improve programs. Human developers do not expect the first version of their code to be perfect. They run the program, examine the results, correct errors, apply different algorithms, and compare whether performance improves. ERA automates this process of trial and error while allowing far more candidates to be tested rapidly than a human could examine.

An important condition is that the research problem must be measurable through a score. If the quality of a program can be quantified through measures such as predictive accuracy, processing speed, error rate, or the degree of data integration, ERA can search for code that improves those scores. Once the researcher provides the problem description, data, and evaluation criteria, the system repeatedly cycles through generating, executing, and modifying candidate solutions.

In this approach, the evaluation and selection structure is more important than the act of writing code itself. Even when the language model produces plausible-looking code, candidates with poor actual performance are eliminated during the search process. Conversely, a method that initially appears ordinary may become the basis for further variations if it performs well in evaluation. Unlike a conventional conversational AI, ERA?™s next action is determined not by how persuasive an answer sounds but by the outcome of actually executing the program.

ERA transforms scientific software development into a form of evolutionary process. Multiple candidate programs are generated, evaluated in real environments, and the stronger candidates become the starting points for the next generation. The system moves closer to higher scores by replacing parts of a program, adding new computational methods, and combining existing algorithms. In effect, AI carries out at scale the process in which humans would normally think of and test ideas one at a time.

This does not mean that ERA understands the entirety of science and begins research from nothing. Researchers must first define the problem in concrete terms. They must determine which data to use, what should be predicted or classified, and which standards should be used to judge a good result. ERA is strongest when exploring possible implementations after a clear objective and evaluation criteria have been established. Its capability lies less in creating questions than in translating questions into executable calculations and optimizing them.

The ERA study attracted attention because it was applied not merely to coding exercises but to difficult problems in actual scientific fields. The researchers examined its potential across tasks with very different characteristics, including single-cell bioinformatics, infectious disease forecasting, neuroscience, geospatial analysis, time-series forecasting, and numerical computation.

A representative case was single-cell RNA analysis. Even when cells are collected from the same type of tissue, differences can arise depending on the time of the experiment, the equipment used, the research institution, or the processing method. This is known as the batch effect. Before comparing data obtained from different experiments, researchers must reduce these technical differences while preserving biologically meaningful signals. If batch effects are removed too aggressively, real cellular differences may disappear. If they are not removed sufficiently, differences in experimental conditions may be mistakenly interpreted as biological discoveries.

ERA generated and improved multiple analytical programs designed to address this problem. In the researchers??evaluation, many of the methods created by ERA scored higher than leading human-developed methods in public performance benchmarks at the time. The researchers reported 40 candidate methods that surpassed existing top-performing approaches in single-cell data integration tasks. This suggests that AI can do more than simply replicate existing analytical procedures. It can produce competitive results by newly combining and modifying multiple methods.

The system also delivered meaningful results in forecasting infectious disease hospitalizations. During the COVID-19 pandemic, research institutions around the world developed various models to predict how many patients would be hospitalized in the coming weeks. These forecasts could be used to prepare hospital beds, medical personnel, and medicines, but the course of an infectious disease is affected by numerous variables, including new variants, seasons, vaccination rates, and changes in human behavior.

ERA repeatedly generated and improved programs for forecasting COVID-19 hospital admissions. The researchers explained that 14 models developed by ERA outperformed the U.S. Centers for Disease Control and Prevention ensemble and the individual forecasting models used for comparison. An ensemble combines multiple prediction models to reduce errors, so surpassing it suggests that the code and model structures explored by ERA moved beyond simple automated code generation.

The study also included a task involving the prediction of neural activity in zebrafish. The researchers used ERA to develop software capable of predicting complex activity patterns observed across tens of thousands of neurons. Neural data are both large in scale and constantly changing over time, which means that model structure and computational efficiency are equally important. ERA demonstrated the possibility that performance could also be improved in problems requiring both biological background knowledge and large-scale computation.

Its results across multiple fields show that ERA is not a dedicated program customized for only one task. The central issue is not whether the system perfectly understands each scientific discipline. If a problem can be expressed as an executable software task and a score can be provided to measure program quality, the same search structure can be applied to a wide range of research fields.

A high score, however, does not automatically constitute a scientific discovery. Even when a program outperforms existing methods on a benchmark, researchers must confirm whether that performance continues on new data and whether the program has been overfitted to specific conditions. They must also scientifically interpret why the program produced good results. Software created by ERA may be a powerful tool for accelerating discovery, but it does not automatically guarantee verification or interpretation.

The claim that AI can produce scientific software at the level of human experts can easily lead to the expectation that the role of scientists will diminish. A closer look at the way ERA works, however, makes the role of human researchers even clearer.

For ERA to operate, someone must first define the scientific problem. Determining which phenomenon needs to be explained, which data can be trusted, what must be predicted, and which result should be considered good remains the responsibility of the researcher. The system concentrates on maximizing a given score, but it has difficulty judging on its own whether that score is scientifically appropriate.

For example, when a model is optimized only for the accuracy of infectious disease forecasts, the selected model may become excessively fitted to data from a particular region or period. Its average score may be high while it fails to reflect risks faced by older people or vulnerable regions. In bioinformatics, maximizing a data-integration score may also dilute biologically important differences. When evaluation criteria are poorly designed, AI may create a program optimized for test performance rather than one that is scientifically meaningful.

Data leakage is another important concern. If information that should be reserved for validation becomes mixed into the program-development process, performance can appear higher than it truly is. Even when researchers do not intend it, ERA may adapt too closely to the characteristics of the evaluation data while testing large numbers of code variations. Training and validation data must therefore be strictly separated, and the resulting software must be tested again on completely independent datasets.

Code stability and reproducibility must also be examined. A program is difficult to trust as a research tool if it operates only in a particular computing environment or if its results vary depending on the version of an external library. Researchers must verify whether the same data and conditions produce the same results, whether the calculation process is properly recorded, and whether other scientists can understand and validate the code.

The provenance of AI-generated programs also remains an issue. If a new method is built by drawing on existing papers and open-source code, researchers should be able to trace where its ideas originated. Questions about who deserves credit for the research outcome and who is responsible when the program contains an error will also require further discussion.

These limitations do not mean that ERA has little value. Rather, they mean that as AI greatly increases the speed of implementation, humans must devote more attention to problem definition and verification. The central role of the scientist may shift from writing every line of code personally to asking good questions, designing evaluation criteria, interpreting results, and examining possible sources of error.

Researchers must become supervisors who determine the direction of exploration rather than passive users who simply accept AI-generated results. When an unexpected result appears, they must distinguish whether it signals a new discovery, a data error, or a weakness in the evaluation criteria. They must also explain what scientific meaning lies behind a high-performing program. Automation of implementation does not eliminate human judgment; it changes where judgment is most needed.

The greatest significance of ERA does not lie only in the fact that a single AI system produced better code than humans. Its more important implication is that the time structure of scientific research itself may change.

In conventional research, implementing a single analytical idea could take days or weeks. Comparing multiple methods required even more time. When a research team lacked sufficient personnel or funding, it was difficult to test every promising possibility. Researchers had to select only a few approaches that appeared most likely to succeed, while the rest were abandoned because of time constraints.

Systems like ERA can reduce the cost of exploration. Once a researcher provides the problem and evaluation criteria, AI can create various algorithms and code structures, execute many candidates in parallel, and continue exploring around the most promising approaches. If analytical candidates that once required a month of human testing can be compared in a much shorter period, scientists will be able to test more ideas against actual data.

This matters not only for large research institutions but also for small research teams. Researchers who cannot secure specialized scientific software developers may gain a better chance to implement their hypotheses. Customized analytical tools may also be developed in areas that are important but have attracted limited markets and research personnel, such as rare diseases or local environmental problems.

The rate at which scientific software is improved may also accelerate. Until now, many research programs were developed for a specific paper and then left without maintenance or further development. When the original developer moved to another laboratory or research funding ended, errors went uncorrected and new data formats were not supported. If ERA is used to evaluate and improve existing programs, it may become possible to update older scientific tools for modern environments and increase their computational efficiency.

This change, however, does not automatically guarantee the democratization of science. Operating systems like ERA requires substantial computing resources, data, and specialized validation frameworks. If institutions with high-performance AI and computing infrastructure can explore more programs more quickly, the research gap may widen rather than shrink. A new divide may emerge between scientists who can access these systems and those who cannot.

There is also a risk that research will become excessively focused on problems that are easy to evaluate. ERA is strongest when program quality can be judged with a clear score. Not every scientific question, however, can be reduced to a single number. It is difficult to establish one evaluation criterion for theoretical research that creates new concepts, studies that interpret hard-to-observe phenomena, or early exploratory work in which it is not yet clear which variables should even be measured.

If only problems that are easy to score are selected, science may become overly focused on measurable performance competition. Tasks that produce strong benchmark scores in the short term may receive priority over research with greater social value or long-term significance. ERA should therefore be used not as a system that selects scientific questions on behalf of humans, but as an infrastructure that helps researchers test questions they have judged to be important more rapidly.

Research culture must also change. AI-generated programs and human-written programs should be evaluated under the same standards, and their generation processes and evaluation data should be disclosed transparently. Researchers should record which models and computing resources were used, how many candidate programs were tested, and whether performance improvements were reproduced on independent data. Without such practices, scientific research may become filled with software that produces excellent results but follows processes that cannot be verified.

The arrival of ERA is less like an event in which AI takes the place of scientists and more like the introduction of a new form of experimental equipment. Just as microscopes opened a world invisible to the human eye and telescopes extended observation into the universe, AI-based software-search systems expand the range of analytical methods humans can test within limited time.

Scientists have long spent enormous amounts of time translating ideas into code. In the future, while AI rapidly generates and improves implementation candidates, researchers may concentrate more on the meaning of the question, the evaluation criteria, and the interpretation of the results. The basis of research competitiveness may also shift from how quickly someone can personally write code to which problems they select, how they define a good result, and how rigorously they verify the answers produced by AI.

ERA will not solve every scientific problem. It cannot determine the purpose of research on its own, nor can it perfectly distinguish between a high score and an important discovery. It does, however, demonstrate the possibility of turning into reality research that could not previously be attempted because there was not enough time or personnel to implement an idea. If the speed of science depends not only on the number of ideas but also on the ability to execute them, ERA is a new tool that expands that capacity for execution.

The fact that AI has begun to improve scientific software on its own is not a declaration that machines have become scientists. It is a sign that scientists can explore a greater number of methods and that the time between hypothesis and verification is beginning to shrink. The central competition in science may no longer be about who can write the most code personally, but about who can select the most important questions, design the most accurate evaluation criteria, and build the most reliable systems of verification.

Nature, May 2026, Aygun, E., et al., An AI System to Help Scientists Write Expert-Level Empirical Software