What It Means for AI to Participate in Discovery

Scientific discovery has always been a slow, human process.

For most of modern history, discovery has followed a specific rhythm: observation → hypothesis → interpretation. Progress is uneven, shaped as much by judgement and imagination as by data.

This rhythm hasn't disappeared, but it has changed subtly over time. Computation is no longer confined to analyzing results after the fact. In some domains, it is beginning to shape the discoveries themselves.

AI is beginning to participate in discovery.

This is not to say AI is replacing scientists… far from it. Rather, it is helping expand what can be explored.

Let me explain.

For much of its early life, AI helped scientists process data they already had. It detected patterns and optimized known processes. Today, in some contexts, it is increasingly used upstream, generating hypotheses that help humans reach places we previously could not.

This doesn't mean machines are “doing science.” It doesn't imply general intelligence, agency, or intention. And it especially does not replace human judgement.

Meaning still comes from people. Responsibility still rests with people. What has changed is the terrain we are able to traverse.

This shift became undeniable when AI started feeling like a thinking partner.

I first noticed this with the introduction of ChatGPT. It was the first time I experienced a system that could engage closely with my thinking… not by asserting answers, but by helping me explore questions, assumptions, and possibilities in real-time.

The system was not intelligent in a human sense, and it was certainly not authoritative. But it changed how ideas could be developed, tested, and refined. It didn’t tell me what was true. It changed what felt feasible to think through.

That difference, while subtle, carries real weight.

The roots of this shift can be traced back to the physics-inspired foundations of machine learning.

John Hopfield and Geoffrey Hinton. Illustration by Niklas Elmehed. © Nobel Prize Outreach.

Some of the most important foundations of modern machine learning did not come from attempts to imitate human cognition, but from physics.

As a testament to this impact, this year’s Nobel Prize in Physics was awarded to two pioneers of machine learning: John Hopfield and Geoffrey E. Hinton, “for foundational discoveries and inventions that enable machine learning with artificial neural networks.”

John Hopfield, a professor at Princeton University, invented the Hopfield Network, an associative memory system that can store and reconstruct patterns, like images, from data. Meanwhile, Geoffrey Hinton, often called the “godfather of deep learning,” developed the Boltzmann Machine, a type of neural network that can autonomously discover properties in data. This enables it to perform tasks such as identifying specific elements in pictures. These innovations laid the groundwork for the AI systems we rely on today, and both are rooted in the principles of physics.

Hopfield and Hinton’s discoveries show how physics itself inspired the development of AI, leading to innovations that now help us push the boundaries of both science and technology.

Rooted in physical principles, their work laid essential groundwork for the machine learning methods used across science today.

“The laureates’ work has already had a tremendous impact. In physics, we use artificial neural networks in a vast range of areas, from developing new materials to identifying specific properties.”
— Ellen Moons, Chair of the Nobel Committee for Physics

In chemistry and biology, computational models have reshaped what discovery makes possible.

David Baker, Demis Hassabis and John Jumper. Illustration by Niklas Elmehed. © Nobel Prize Outreach.

A similar pattern appears in the life sciences.

The Nobel Prize in Chemistry was awarded in 2024 to David Baker, Demis Hassabis, and John Jumper for breakthroughs in computational protein design and protein structure prediction.

David Baker, a professor at the University of Washington and investigator at Howard Hughes Medical Institute, was able to use different amino acids to design a new kinds of proteins. On the other hand, Demis Hassabis and John M. Jumperfrom Google DeepMind developed AlphaFold2, a model that can predict the 3D structure of proteins. In a groundbreaking achievement, AlphaFold2 has predicted the structures of nearly all known proteins—around 200 million of them.

This work did not remove the need for experiments or biological understanding. Instead, it fulfilled a decades-old scientific hope by making an otherwise intractable problem navigable.

Human interpretation, validation, and responsibility remain central, but the field of view has expanded dramatically.

“These discoveries fulfill a 50-year-old dream of predicting protein structures and open up vast possibilities for future research and applications in medicine, biotechnology, and beyond.”
— Heiner Linke, Chair of the Nobel Committee for Chemistry

As computational tools grow more powerful, human judgment becomes more important, not less.

Across these examples, a clear pattern emerges.

The need for discernment, responsibility, and care has not changed. What has changed is the scale and speed at which scientific possibilities can be generated.

When models surface vast numbers of plausible candidates, deciding what matters becomes harder, not easier. Power concentrates responsibility rather than dissolving it.

Science and technology are human activities, and their development must remain ordered toward the good of the human person.

Artificial intelligence, used well, does not replace discovery. It changes how discovery unfolds. And that change asks something serious of us: patience, clarity, and humility equal to the tools we now wield.