AI for science needs reasoning, not just data
2026-08-20 · MIT Technology Review
AI for Science Needs Reasoning, Not Just Data
The Recurring Announcement That Science Has Ended
Every few decades, someone announces that science has reached its end. In 1903, physicist Albert Michelson wrote that the “facts of physical science have all been discovered.” In the 1980s, Stephen Hawking predicted that theoretical physics might be finished by the end of the century. With the explosive arrival of artificial intelligence, the feeling is in the air again—this time accompanied by a Nobel Prize.
AlphaFold’s Nobel-Winning Breakthrough
In 2024, Demis Hassabis and John Jumper of Google DeepMind were awarded part of the Nobel Prize in Chemistry for their neural network AlphaFold. AlphaFold predicts the three-dimensional structures of proteins by learning from thousands of experimentally measured shapes. This difficult problem had resisted systematic attacks for half a century. AlphaFold seemed to have solved it once and for all.
Hassabis and his team called AlphaFold “the template for how AI can accelerate all of science to digital speed.” A wave of startups building foundation models for biology, chemistry, and materials discovery raised billions of dollars, inspired by DeepMind’s success. AlphaFold demonstrated that the combination of AI and sufficient data could make groundbreaking discoveries, even without understanding the underlying mechanisms. It appeared that a clear path through the rest of science had been laid out.
Why AlphaFold’s Conditions Are Rare
Although AlphaFold is a profound achievement, the conditions that enabled it are uncommon, and replicating them in other fields will take decades rather than years. The primary condition for its success was the existence of the Protein Data Bank, a dataset of roughly 170,000 experimentally validated protein structures.
Creating the Protein Data Bank was not simple. It required 53 years of international scientific cooperation and roughly $21 billion worth of experimental work. Efforts of that scale are notoriously difficult to fund, nearly impossible to coordinate, and extremely time-consuming; many have failed as a result.
Barriers Beyond Data Availability
Even in fields with sufficient cohesion and resources where data is not locked behind commercial ownership, another barrier is too little discussed: the scientific impossibility of generating comparable data at scale.
The key experimental technique for protein structures—protein crystallography—is unusually replicable and dependable, supporting more than 25 Nobel Prizes. However, in most experimental science, results vary more often than not. Cell lines drift. Chemicals have trace contaminants. Lab humidity changes.
Creating measured datasets consistent enough, accurate enough, and scalable enough to train modern neural networks in biology or most of chemistry would require new kinds of measurement and standardized approaches that will not be ready anytime soon.
Fields Where Data-Driven Breakthroughs Remain Possible
A handful of fields already meet these stringent requirements: weather forecasting, much of genomics, and very limited areas of chemistry. These domains may see AlphaFold-style breakthroughs soon. Government support for producing and coordinating high-quality datasets will be critical, as argued by the US National Security Commission on Emerging Biotechnology.
For most open questions in science, however, a different plan is needed, at least in the short term.
The Rise of AI Agents
Scientists have always reasoned under uncertainty. Biologists identifying new drug targets never had perfect datasets. They combine docking calculations, known structures, molecular dynamics, binding assays, and use judgment to weigh each method’s strengths and weaknesses. The real skill of science lies in synthesizing multiple imperfect tools and revising conclusions as evidence arrives.
Until recently, no software could perform this synthesis. AI agents now can.
An agent is an AI reasoning engine given access to tools—digital or physical—and the ability to use them. Recent architectural advances powered by large language models have enabled these systems, dramatically reducing the need for scientifically specialized datasets.
For science, this is foundational: agents can digitally model the iterative, highly contingent process of actual research. While tools like AlphaFold apply powerful methods to narrow questions, agents are generalists that emulate the human process of discovery rather than replace it.
Google’s AI Co-Scientist: A Practical Example
Consider Google’s AI Co-Scientist, announced in May. Researchers provided it with a one-page brief and a goal: determine how antibiotic resistance spreads between bacterial species, a major driver of drug-resistant infections.
The system created sub-agents with specialized roles:
- One drafted hypotheses from the literature
- Another critiqued them 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 were hitchhiking on bacterial viruses, using whichever virus could ferry them into a new host. The hypothesis was correct. Researchers at Imperial College London had spent a decade arriving at the same insight.
This example illustrates how AI agents can tackle complex scientific problems by reasoning under uncertainty, offering a more general and immediately practical route to accelerating science across disciplines without waiting decades for perfect datasets.