Science Is Becoming Less Human

Updated at 9:36 a.m. ET on December 12, 2023

This summer, a pill intended to treat a chronic, incurable lung disease entered mid-phase human trials. Previous studies have demonstrated that the drug is safe to swallow, although whether it will improve symptoms of the painful fibrosis that it targets remains unknown; this is what the current trial will determine, perhaps by next year. Such a tentative advance would hardly be newsworthy, except for a wrinkle in the medicine’s genesis: It is likely the first drug fully designed by artificial intelligence to come this far in the development pipeline.

The pill’s maker, the biotech company Insilico Medicine, used hundreds of AI models to discover both a new target in the body that could treat the fibrosis and which molecules might be synthesized for the drug itself. Those programs allowed Insilico to go from scratch to putting this drug through the first phase of human trials in two and a half years, rather than the typical five or so. Even if the pill proves useless, a real possibility, plenty of other drugs designed with the help of AI are in the wings. Scientists and companies alike hope that these will reach pharmacies far faster than traditionally designed medicine—bringing a drug to market typically takes well over a decade, and failure rates are high.

Medicine is just one aspect of a broader transformation in science. In only the past few months, AI has appeared to predict tropical storms with similar accuracy and much more speed than conventional models; Meta has released a model that can analyze brain scans to reproduce what a person is looking at; Google recently used AI to propose millions of new materials that could enhance supercomputers, electric vehicles, and more. Just as the technology has blurred the line between human-created and computer-generated text and images—upending how people work, learn, and socialize—AI tools are accelerating and refashioning some of the basic elements of science. “We can really make discoveries that would not be possible without the use of AI,” Marinka Zitnik, a biomedical and AI researcher at Harvard, told me.

Science has never been faster than it is today. But the introduction of AI is also, in some ways, making science less human. For centuries, knowledge of the world has been rooted in observing and explaining it. Many of today’s AI models twist this endeavor, providing answers without justifications and leading scientists to study their own algorithms as much as they study nature. In doing so, AI may be challenging the very nature of discovery.

AI exists to derive impossibly intricate patterns from data sets that are too large for any person to fathom, a mystifying phenomenon that has grown more familiar since ChatGPT was released last year. The chatbot—a tool suddenly at everyone’s fingertips that appears to synthesize the entire internet—changed how we can access and apply knowledge, but it simultaneously tainted much of our thinking with doubt. We do not understand exactly how generative-AI chatbots determine their responses, only that they sound remarkably human, making it hard to parse what is real, logical, or trustworthy, and whether writing, even our own, is fully human or bears a silicon touch. When a response does make sense, it can seem to offer a shortcut rather than any true understanding of how or why the answer came to be.

AI may be doing something similar in a broad range of scientific disciplines. Among the most notable scientific advances achieved via AI may be those in molecular biology from DeepMind, a leading AI research lab now based at Google. After DeepMind’s programs conquered the game of Go in 2016—a game so much more complex than chess that many thought that computers could never master it—Demis Hassabis, DeepMind’s CEO, told me that he began considering how to build an AI program for the decades-old challenge of protein folding. All sorts of biological processes depend on proteins, and every protein is made of a sequence of amino acids. How those molecules fold into a three-dimensional shape determines a protein’s function, and mapping those structures could help scientists develop new vaccines, kill antibiotics-resistant bacteria, and explore new cancer treatments. Without a protein’s 3-D shape, scientists have little more than a bunch of Lego bricks without instructions for putting them together.

Figuring out a single protein structure from a sequence of amino acids used to take years. But in 2022, DeepMind’s flagship scientific model, AlphaFold, found the most likely structure of almost every protein known to science—some 200 million of them. Much like the company’s chess- and Go-playing programs, which search for the best possible move, AlphaFold searches the number of possible structures for an amino-acid sequence to find the most probable one. The program compresses what could have been an entire Ph.D.’s worth of work into seconds, and it has been widely lauded for its “revolutionary impact” on basic biology and the development of novel treatments alike. Still, independent researchers have noted that despite inhuman speed, the model does not fully explain why a specific structure is likely. As a result, scientists are trying to demystify AlphaFold’s predictions, and Hassabis noted that those efforts are making good progress.

Read: Welcome to the big blur

AI allows researchers to study complex systems in “the world of bits” at a much faster pace than in the “world of atoms,” Hassabis said, and then physically test their hypotheses as a final step. The technology is pushing forward advances in numerous other disciplines—not just improving speed and scale, but changing what kind of research is thought possible. Neuroscientists at Meta and elsewhere, for instance, are turning artificial neural networks trained to “see” photographs or “read” text into hypotheses for how the brain processes both images and language. Biologists are using AI trained on genetic data to study rare diseases, improve immunotherapies, and better understand SARS-CoV-2 variants of concern. “Now we have viable hypotheses, where before we had mysteries,” Jim DiCarlo, a neuroscientist at MIT who has pioneered the use of AI to study vision in the brain, told me.

Astronomers and physicists are using machine learning to process data sets from the universe that were too immense to touch before, Brice Ménard, an astrophysicist at Johns Hopkins, told me. Some experiments, such as the CERN particle collider, produce too much information to physically store. Researchers rely on AI to throw out familiar observations while keeping unknowns for analysis. “We don’t know what the needle looks like, because these are undetected physics events, but we know what the hay looks like,” Alexander Szalay, the director of the Institute for Data Intensive Science at Johns Hopkins, told me. “So computers are trained to recognize the hay and basically throw it away.”

Read: Computers are learning to smell

The long-term vision could even involve combining AI models and physical experiments in a sort of “self-driving lab,” Zitnik said, wherein computer programs and robots generate hypotheses, plan experiments to test them, and analyze the results. Such labs are a ways off, although prototypes do exist, such as the Scientific Autonomous Reasoning Agent, a robotic system that has already discovered new materials for renewable energy. SARA uses a laser to analyze and alter materials iteratively, with each loop lasting a few seconds, Carla Gomes, a computer scientist at Cornell, told me—reducing days of research to hours. This future, if it comes to pass, will elevate software and robots from tools to collaborators, even co-creators of knowledge.

Quantum observations too numerous for humans to store, experiments too rapid for humans to run, neuroscientific hypotheses too complex for humans to derive—even as AI enables scientific work never before thought possible, those same tools pose an epistemic dilemma. They will produce groundbreaking knowledge while breaking apart what it means to know in the first place.

“The holy grail of science is understanding,” Zitnik said. “To be able to understand a phenomenon, whether that’s the behavior of a cell or a planetary system, requires being able to identify causes and effects.” But AI models are famously opaque. They detect patterns based on gargantuan data sets via software architectures whose inner workings baffle human intuition and reasoning. Experts have taken to calling them “black boxes.”

This presents obvious problems for the scientific method. “We have to understand what is going on inside this black box so we can see where this discovery is coming from,” Szalay told me. To predict events without understanding why those predictions are accurate might gesture toward a different type of science, in which knowledge and the resulting actions are not always accompanied by an explanation. An AI model might predict a thunderstorm’s arrival but struggle to explain the underlying physics and atmospheric changes that triggered it, analyze an X-ray without showing how it arrived at its diagnosis, or propose abstract mathematical conjectures without proving them. Such shifts from observations and grounded reasoning to mathematical probability have happened in the sciences before: The equations of quantum mechanics, which emerged in the 20th century, accurately predict subatomic phenomena that physicists still don’t fully understand—leading Albert Einstein himself to doubt quantum theory.

From the January 1951 issue: Faith in science

Science itself may offer a solution to this conundrum. Physical experiments have uncovered a great deal in the past century about the quantum world, and similarly, AI tools might appear inscrutable partly because researchers haven’t spent enough time probing them. “You have to build the artifact first before you can pull it apart and scientifically analyze it,” Hassabis told me, and scientists have only recently begun to build AI models worthy of study. Even older numerical simulations, although far less complex than today’s AI models, are hard to interpret in an intuitive way, but they have nevertheless informed new discoveries for decades.

If researchers understand how artificial neurons respond to an image, they might be able to translate those predictions to biological neurons; if researchers understand which parts of an AI model link a mutation to a disease, scientists could gain new insights into the human genome. Such models are “fully observable systems. You can measure all the parts,” DiCarlo said. Whereas he cannot measure every neuron and synapse in a monkey brain during surgery, he can do that for an AI model. With the right access, AI programs might present scientists not with black boxes so much as with a new type of object requiring a new sort of inquiry—not “models” of the natural world so much as addenda to it. Some scientists even hope to build “digital twins” to simulate cells, organs, and the planet.

AI is not a silver bullet, though. AlphaFold may be revolutionary, and perhaps Insilico will indeed drastically reduce the time it takes to develop new medicine. But the technology has significant limitations. For instance, AI models need to train on large amounts of relevant data. AlphaFold is a “spectacular success,” Jennifer Listgarten, a computational biologist and computer scientist at UC Berkeley, told me—but it also “relied on very expensive, highly curated data that was generated over decades in the laboratory on a very crisply defined problem that could be evaluated extremely cleanly.” The lack of high-quality data in other disciplines can prevent or limit the use of AI.

Even with those data, the real world can be more complex and dynamic than a silicon simulation. Translating the static structure of a molecule into its interactions with various systems in the body, for instance, is a problem that researchers are still working on, Andreas Bender, who studies molecular informatics at the University of Cambridge, told me. AI can propose new medicines quickly, but “you still need to run the drug-discovery process, which is, of course, quite long,” John Jumper, a researcher at DeepMind who led the development of AlphaFold, told me.

Clinical trials take years and many are unsuccessful; plenty of AI drug start-ups and initiatives have scaled back. Those failures are, in some sense, evidence of science working. Experimental results, along with known physical laws, allow scientists to prevent their models from hallucinating, Anima Anandkumar, a computer scientist at Caltech, told me. No analogous laws of linguistic accuracy exist for chatbots—consumers have to trust Big Tech.

In a lab, novel predictions can be physically and safely tested in an isolated setting. But when developing drugs or treating patients, the stakes are much higher. Existing maps of the human genome, for instance, skew toward white Europeans, but the expression of many conditions, such as diabetes, varies significantly by race and ethnicity. Just as biased data sets produce racist chatbots, skewed biological data might mean that “models are not applicable to people of non-Caucasian origin,” Bender told me, or those of different ages, or with existing diseases and on co-medications. A cancer-diagnosis program or treatment designed by AI might be especially effective only on a small slice of the population.

AI models might transform not just how we understand the world but how we understand understanding itself. If so, we must build out new models of knowledge as well—what we can trust, why, and when. Otherwise, our reliance on a chatbot, drug-discovery tool, or AI hurricane forecast might depart from the realm of science. It might be more akin to faith.

This article previously misstated Jim DiCarlo’s title.