AI Is Changing Science, but Nobel Laureates Say Trust Still Comes From Old-Fashioned Testing

In today’s world, it seems like AI is everywhere. It writes emails and code, summarizes meetings, analyzes images, acts as a personal assistant, and much more. Scientists sometimes use it not only to process data, but to decide what experiments to try in the first place.

At the 75^th Lindau Nobel Laureate Meeting, this led to a sharper question. AI is obviously powerful, and it clearly has many applications. But what happens when AI becomes a part of the scientific method? Can we trust it, especially when it makes a prediction before humans can understand why?

Can AI Understand Science?

That question was addressed head-on at a panel moderated by Heiner Linke, professor at Lund University.

“Who needs to understand something for it to be understood? Is it humans, is it a model, or how should that work?” Linke asked at the start of the panel.

Anne L’Huillier, who shared the 2023 Nobel Prize in Physics for work on attosecond pulses of light, says understanding is not always a straightforward thing. In the early days of her field, researchers saw something unexpected when atoms were exposed to intense laser fields: high-order harmonics, bursts of light at much higher frequencies than expected. Simulations eventually reproduced the effect. But for L’Huillier, that was not yet true understanding.

“Do we understand the phenomenon? I would say no,” she said. “Understanding is something else. It’s really an intuition and an image.”

The real advance came when physicists developed a model for what was happening. It was a simplified model, of course, but useful enough to let scientists picture attosecond pulses and then pursue them in the lab.

William Phillips, who shared the 1997 Nobel Prize in Physics for laser cooling and trapping atoms, took the thought further. He asked the audience to look closely at his tie, which had a pattern of the familiar Bohr atom, showing electrons orbiting a nucleus like planets around the sun.

Physicists know that image is incorrect. Electrons are described by quantum wave functions, not tiny planetary paths. Yet the image remains everywhere because it is still useful and does something science needs. It gives people a way to think. Phillips was also skeptical that machines would replace the most unruly part of science: surprise.

“I don’t think AI is ever going to discover something that’s unexpected… but who knows,” he said. “The world is full of unsolved problems, there’s plenty that we still have not understood, and plenty of things to be unexpectedly discovered in the laboratory.”

But there is certainly a re-evaluation of what AI can do in science.

Klara Bonneau, a Young Scientist who works on neural-network models for protein dynamics, described a generational shift in how AI is perceived. During her early studies, she said, AI was often treated as a shortcut for people who did not want to understand the underlying science. Now, AI is used in groundbreaking research.

It was very fitting, then, that the also directed the panel to think about trust. Historically, trust in science is linked to reproducibility and traceability. Now, with AI often working as a black box, that is broken.

“So do you think we can, or should we learn to trust something despite this lack of prediction traceability?” Bonneau asked.

AI and Proteins

Few people are more qualified to answer that question than John Jumper. Jumper, who shared the 2024 Nobel Prize in Chemistry for AlphaFold, argued that science has faced versions of this problem before. Simulations can produce predictions without giving scientists a simple explanation. He gave the example of crystallographers who may find the conditions that make a protein crystallize without fully understanding why those conditions worked.

In his view, AI is another object scientists can interrogate. It’s not the end of understanding, but rather a tool, another stage in the scientific discovery process. For Jumper, trust in AI begins with good old-fashioned testing. You need to know what was in the training set, what was held out, and whether the test data resembled the way real scientists would use the system. He also argued for testing models on future data, because AI is trained on the past and then used on problems that come later.

Linke, who serves as the chair of the Nobel Committee for Chemistry, further emphasized that AlphaFold was Nobel-worthy because it met the standards of a real scientific breakthrough.

Firstly, AlphaFold was published so that others could inspect and reproduce it. AlphaFold was trained on the International Protein Data Bank, which is publicly accessible and curated by scientists, so it’s not just a random scraped dataset. Lastly, the predictions were verifiable. AlphaFold’s database now provides open access to more than 200 million protein structure predictions.

Jumper also pushed back against the idea that AlphaFold can only reproduce what it has already seen. He said people claim it cannot predict a “novel fold” because that is novel. But as we’ve seen since the already famous AlphaGo “Move 37”, AI can apparently create novel ideas. The same thing happened with AlphaFold, where people claimed it wouldn’t be able to predict a “novel fold” because that is novel. But in tests, proteins labeled as novel folds were predicted about as accurately as other proteins. Jumper continued by saying AI can also help in a different way: by helping researchers ask good questions.

How Laureates and Young Scientists Are Using AI

Outside the panel, AI was permeating throughout the Lindau Meeting. Omar Yaghi was jointly awarded the 2025 Nobel Prize in Chemistry. Yaghi builds materials from molecular building blocks, joining metal nodes and organic linkers into porous structures known as metal-organic frameworks, or MOFs, and related covalent organic frameworks, or COFs.

A single gram of a highly porous MOF, he said, can contain almost the surface area of a football field. By changing pores and chemical groups, researchers can design materials that bind carbon dioxide, take up water from dry air or separate molecules from mixtures. But there are far too many possible materials for humans to explore by hand.

“So we’ve gone from a world of scarcity,” he said, “to the world of abundance.” That abundance changes the role of the chemist. The hard part isn’t making a new material, but deciding what are the best materials to make. He said students in his lab use commercially available AI, feeding them information and then asking the system to return a list of suggestions. Many such suggestions are obvious, but some are actually useful. After three cycles, Yaghi said, the group obtained materials “three times more crystalline than what is reported” in two weeks rather than years.

Meanwhile, Ferenc Krausz who shared the 2023 Nobel Prize in Physics for attosecond science, offered a very different example from medicine. He wants to analyze blood in detail and then use AI algorithms to analyze that data and find signs of diseases and conditions before any symptoms become apparent.

Among young researchers, AI also got a great deal of attention. Alex Plum, a biophysics researcher at the University of California, San Diego, is treating embryogenesis as an information problem. How does a developing tissue preserve, lose or transform positional information while cells are moving? In that framing, an embryo is not just growing. It is processing information through time.

Mykyta Kliapets, working in astronomy, is using machine learning for a different kind of application: the torrents of data from modern sky surveys. Stellar light curves can reveal whether stars are pulsing, rotating, eclipsing or distorted by instrumental noise. But the next generation of surveys will produce too many observations for old approaches to keep up.

Larissa Fischer’s work turns to the brain before dementia is visible. Her published work on cognitively unimpaired older adults suggests the early story is not simple, but AI can help make more sense of this data.

The Science (and Art) of Modelling Reality

Meanwhile, Pat Hanrahan’s Heidelberg Lecture offered a different route into the same problem. The Heidelberg Laureate Forum is a “cousin” of the Lindau Nobel Laureate Meetings, celebrating laureates in the fields of mathematics and computer science. Hanrahan, who received the 2019 ACM A.M. Turing Award, is also celebrated for his work in movies like Toy Story.

Before AI became the hot topic it is today, Hanrahan was trying to make computers model the visible world. His career helped build modern computer graphics, including the tools that made full-length computer-generated films possible.

He described the enormously difficult task of turning messy reality into a believable computation. You need models of mountains, grass, roads, rain, atmosphere, surfaces, light and materials. You need enough physics to make the result believable, and enough computation to make it possible.

His Lecture highlighted ordinary things that are surprisingly hard to simulate. Skin, for instance, doesn’t look real if treated as a simple surface. Much of what we see as skin color comes from light entering the tissue, scattering below the surface and reflecting from blood and other layers. To simulate this, you need an elaborate physical model and billions of polygons.

Then there is the always-important problem of light. In a room, much of what we see is indirect illumination: light bouncing from windows to walls, from walls to objects, from objects back to the eye. To simulate that, computer graphics researchers used physically based rendering and path tracing, following the possible paths light can take through a scene.

Hanrahan’s work was used in movies from Toy Story to Terminator II, and in addition to his research accolades, he has also received three Academy Awards for his work in rendering and computer graphics research.

Ultimately, AI is not the first time science has handed part of its imagination to a machine. It started with computers used to solve complex calculations, then we moved on to sifting signals from noisy data, simulating molecules, modelling the climate, and even building imaginary worlds realistic enough for movies.

AI may be faster, stranger, and more autonomous than earlier tools, but the scientific standard should not change. Whether a computer is analyzing a spectrum, rendering a film frame or proposing the next experiment, scientists should still ask the same question: how can we trust this?