AI for Science Needs Reasoning, Not Just Data
Introduction: Has Science Reached Its End?
In an article published by Eric Schmidt, former CEO of Google, Suhas Mahesh, and Maya Levin, they analyze the perception that science has reached its end following the breakthroughs of artificial intelligence. Every few decades, someone declares that science has come to an end. In 1903, the esteemed physicist Albert Michelson wrote that the "facts of physical science have all been discovered." In the 1980s, Stephen Hawking predicted that theoretical physics might reach its end before the close of the century. Now, with the sensational arrival of artificial intelligence, this feeling is hovering in the air again—and this time, it is also accompanied by a Nobel Prize.
In 2024, Demis Hassabis and John Jumper of Google DeepMind won a share of the Nobel Prize in Chemistry for their neural network, AlphaFold. This tool is capable of predicting the three-dimensional structure of proteins by learning from thousands of experimentally measured shapes. This complex problem had posed a tough challenge to systematic attempts to crack it for about half a century; AlphaFold seemed to have solved it once and for all, and the world focused on the promise inherent in this approach. Hassabis and his team called AlphaFold "the template for how AI can accelerate all of science to digital speed." A wave of startups developing foundation models for biology, chemistry, and materials discovery raised billions of dollars, riding on the back of DeepMind's success. The AlphaFold system showed that the combination of AI and a sufficient amount of data could lead to groundbreaking discoveries (even if we do not understand the underlying mechanisms involved), and it seemed, once again, that the path to unlocking the rest of science was paved before us.
The Barriers of the AlphaFold Model: Why Is It Hard to Replicate the Success?
To be sure, AI will bring extraordinary changes to science, but it is becoming increasingly clear that AlphaFold, and tools like it, may not be the best template for that metamorphosis. Although it is a profound achievement, the conditions that enabled the creation of AlphaFold are rare, and the time required to meet these conditions in other fields will be measured in decades, not years. Instead, the acceleration of science will come thanks to another approach: AI agents.
The primary condition for AlphaFold's success was the existence of the Protein Data Bank (PDB), a database of about 170,000 experimentally validated protein structures on which DeepMind’s team could train its model. Creating this database was not simple: it required 53 years of international scientific cooperation and, according to a recent estimate, cost about $21 billion worth of experimental work to assemble. Efforts of this scale are incredibly difficult to fund, almost impossible to coordinate, and highly time-consuming to execute; as a result, they have often been unsuccessful.
But even in fields where the required cohesion and resources exist, and where the relevant data are not blocked due to commercial ownership, there is another barrier that is rarely discussed: the scientific inability to generate comparable data. In the case of protein structures, the key experimental technique—protein crystallography—is an unusually reliable and replicable tool, so much so that more than 25 Nobel Prizes have relied on it. But in most of experimental science, results vary more often than not. Cell lines drift, chemicals have trace contaminants, and lab humidity changes. Creating measured databases that are consistent enough, accurate enough, and scalable enough to train a modern neural network in biology or most fields of chemistry will require new types of measurement and new standardized approaches—and none of these will be ready anytime soon.
Of course, there are a handful of fields where these requirements are met: weather forecasting, much of genomics, and very limited areas of chemistry. These may see AlphaFold-style breakthroughs soon, if they haven't already. Government support for the production and coordination of these databases will be critical, as argued by the US National Security Commission on Emerging Biotechnology. But for most open questions in science, we will need a different plan, at least in the short term. Fortunately, something quieter and more modest has begun to show promise.
The Data Limit and the Transition to AI Agents
Scientists have always reasoned under conditions of uncertainty. Biologists working to identify new drug targets have never held perfect databases. Instead, they combine docking calculations and known structures, factor in molecular dynamics, run a handful of binding assays, and use their judgment to weigh each method according to its strengths and points of failure. The skill of science is not in any single tool; it lies in the synthesis of what many tools produce, and in updating the results as evidence accumulates. This is how most practical research actually proceeds. But until recently, no software could do it.
Agents now can. Simply put, an agent is an AI reasoning engine that has been given access to tools—digital or physical—and the capabilities to use them. Over the last few years, a fundamental architectural shift in AI has enabled the rapid proliferation of these programs, which are powered by large language models (LLMs), while dramatically reducing the need for specialized scientific databases. For science, this technological advancement represents an infrastructural change: it has allowed us to create digital tools that can mimic the iterative, highly contingent process of actual research. While tools like AlphaFold apply a powerful approach to a limited question, agents are inherently generalists. They do not represent a new way to do science—instead, they digitally model the human process of discovery.
Demonstrating Capability: Google's AI Co-Scientist Project
Consider Google’s AI Co-Scientist system, announced in May. Researchers presented it with a one-page brief and a goal: to understand how antibiotic resistance spreads between bacterial species, a key driver of drug-resistant infections. The system spun up sub-agents. One drafted hypotheses from the scientific literature. A second critiqued and broke them down like a peer reviewer. A third ran tournaments to rank the strongest candidates. A fourth refined the winning hypothesis.
The agent concluded that resistance genes hitchhike on bacterial viruses (bacteriophages), using whichever virus can ferry them to a new host. The hypothesis was correct. Researchers at Imperial College London spent a decade reaching the same conclusion through painstaking wet-lab work; their paper, which had not been previously exposed to Co-Scientist, was still in the peer-review stage.
Technological Challenges and Long-Term Impacts
Agents like Co-Scientist are still novel tools, and there are real challenges to overcome before they become an integral part of the scientific process: they are still liable to hallucinate, their judgment is inconsistent, and they have memory and input constraints that limit the time they can run autonomously. But these technical barriers will fall away, and as they do, we will begin to notice the compounding effects of scientific agents on the reliability, consistency, and speed with which science is done.
First, agents offer a structural fix for science's "reproducibility crisis"—the widespread problem of researchers being unable to replicate each other’s results. For decades, the scientific community has begged researchers to share their raw data and exact code in an effort to standardize experimental processes. But researchers have long resisted this tedious administrative work, which happens after the interesting scientific part is already done. Agents, in contrast, automatically log every move they make, creating an exact record of the method that led to their results, allowing for precise replication.
A second consequence will be the amplification of scientific memory. The transfer of knowledge between researchers is a famously murky process; if it isn't done over years of training and observation, graduate students are left to pore through messy lab notebooks kept by predecessors for decades, searching for details that will make or break their protocol. As agents become an increasingly large part of the scientific process, the lab's entire scientific history will be recorded in a central, standardized repository of institutional knowledge.
But the most important impact of agents will be speed. In any field, when testing an idea takes less time than arguing about it in a meeting, people stop debating and just run the test. An agent that can read a thousand papers in an hour, design 500 molecules, and learn from its failed tests by morning will bring down the cost of experimentation and fundamentally change the pace at which science gets done. It will also give researchers the freedom to chase bold, strange questions they never would have risked their time on before, opening scientific doors we have yet to imagine.
The Future of AI Agents in Science
While the AlphaFold template will certainly be key to incredible discoveries, it alone will not bring us to the end of science. Instead, the shift toward agentic AI represents a much rarer tier of breakthrough: a tool that envelops every field of science at once. Historically, tools of such scope have arrived just a handful of times: calculus, statistical inference, spectroscopy, and the computer. Each revealed a world of problems no one had thought to formulate, and those problems, in turn, defined their fields anew. With agents, we face a similar transformation.
Author's Note: Eric Schmidt was the CEO of Google from 2001 to 2011. In 2024, with his wife Wendy, he co-founded Schmidt Sciences, a philanthropic venture to fund unconventional areas of exploration in science and technology. Suhas Mahesh leads AI for Science work at the AI Center of Schmidt Sciences, and he is a specialist in AI for materials discovery. Additional research was conducted by Maya Levin, associate and sciences lead in the Office of Eric Schmidt.