AI for science needs reasoning, not just data
Every few decades, someone announces that science has reached its end. In 1903, the revered 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…
For centuries, scientists have been captivated by the idea that each major breakthrough will herald the end of inquiry in their field. In 1903, Albert Michelson declared the "facts of physical science have all been discovered." The 1980s saw Stephen Hawking predict theoretical physics might reach its end by the century's end. Now, with the advent of artificial intelligence, there is renewed speculation—this time buoyed by a Nobel Prize.
In 2024, Demis Hassabis and John Jumper from Google DeepMind received part of the chemistry Nobel for their neural network AlphaFold, which predicts protein structures by learning from thousands of experimentally measured shapes. This achievement seemed to solve a half-century-old problem, sparking a wave of startups building AI models for biology, chemistry, and materials discovery, fueled by DeepMind’s success.
However, experts now caution that AlphaFold and similar AI models may not be the ultimate blueprint for advancing science. While AlphaFold demonstrated what AI can achieve with sufficient data, its success is rare and relies on extensive, standardized datasets that are not universally available. The creation of such datasets, like the Protein Data Bank containing roughly 170,000 experimentally validated protein structures, took decades of international cooperation and billions of dollars.
Yet, replicating this level of data generation in other fields is challenging due to the inherent variability in biological and chemical systems. In most experimental sciences, results vary more than they converge, making the creation of comparable datasets difficult, if not impossible, with current technologies. This barrier has led some to question the feasibility of AI-driven breakthroughs in most of science in the near future.
However, there is hope on the horizon. Scientists have long relied on reasoning under uncertainty, combining various tools and their judgment to arrive at conclusions. With the recent advancements in AI, digital agents—reasoning engines equipped with tools like digital or physical resources—can now mimic the iterative and contingent process of scientific discovery.
These agents, powered by large language models, can use multiple methods, weigh their strengths and weaknesses, and revise results as new evidence emerges. For example, Google’s AI Co-Scientist was given a brief to understand antibiotic resistance spread among bacterial species. It generated hypotheses, evaluated them, and ran trials, ultimately concluding that resistance genes hitch rides on bacterial viruses.
This approach, while not a new way to conduct research, represents a foundational change: digital tools can now emulate the complex, real-world research process. While tools like AlphaFold tackle specific questions with great precision, agents offer a more generalizable approach, modeling the human research process for a wide range of scientific inquiries.
Government support for data coordination and production will remain crucial, but for most open scientific questions, the key may lie in leveraging AI agents to reason, synthesize, and adapt—mimicking the very essence of scientific inquiry.
Written by urgent.news from MIT Technology Review's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
