Throughout history, major advances in science have often followed the invention of new instruments that expanded what scientists could observe and explore. The telescope opened the heavens to systematic observation. The microscope revealed an invisible world of cells and microorganisms. Computers enabled scientists to model and analyze systems of previously unimaginable complexity.
Artificial intelligence may be the latest — and perhaps the most powerful — addition to this long line of instruments of scientific discovery.
Over the past several years, AI has achieved a number of remarkable scientific breakthroughs, from predicting the three-dimensional structures of proteins to accelerating the discovery of new materials. These advances raise a much larger question: Is AI simply giving scientists a new generation of powerful tools, or is it beginning to change the very process by which scientific discoveries are made?
The Winter/Spring 2026 issue of Dædalus, the journal of the American Academy of Arts & Sciences, is devoted to this question: “What is the future of scientific discovery in this new age of AI?”
Thirty-three scientists responded. “Bringing perspectives from life sciences and medicine, cognitive science and neuroscience, the physical and earth sciences, chemistry and materials science, computer science, mathematics and the social sciences—they draw on their work at the frontier of AI and science.”
James Manyika, Senior Vice President at Google and Chairman Emeritus of the McKinsey Global Institute, served as the issue’s guest editor.
“The combination of progress in AI and its related techniques, its expanding and general usefulness in science, and its celebrated landmark contributions to discovery has led in recent years to recognition by the scientific community that a new era of AI for science has arrived,” Manyika wrote in the issue’s “Introductory Notes.” “Indeed, many universities, research consortia, national academies, national laboratories, and governments have launched AI-for-science initiatives.”
The scientific significance of AI is reflected in several of the world’s most prestigious awards. The 2024 Nobel Prize in Physics went to Geoffrey Hinton and John Hopfield for their pioneering work on machine learning with artificial neural networks. The 2024 Nobel Prize in Chemistry recognized David Baker, Demis Hassabis, and John Jumper for their work on protein design and structure prediction. The 2018 Turing Award — the so-called Nobel Prize of computer science — went to Yoshua Bengio, Geoffrey Hinton, and Yann LeCun for their pioneering work on deep learning. And the 2024 Turing Award recognized Andrew Barto and Richard Sutton for their contributions to reinforcement learning.
These awards illustrate what Manyika calls the “bidirectional nature of progress at the intersection of AI and science: science advancing AI and AI advancing science.”
“To these AI-enabled advances we could add many more from just the last few years in genomics, neuroscience, health care, chemistry, materials science, plasma physics, astrophysics, climate science, and beyond,” he added.
The arrival of AI-enabled science also presents significant challenges. Academic and national laboratories may be poorly equipped for the data- and compute-intensive training required by advanced AI models while facing a continuing brain drain to the private sector, where much frontier AI research and AI-enabled science now takes place.
“Nevertheless,” Manyika wrote, “most scientists acknowledge that AI-enabled science is here to stay and that AI should be harnessed to advance science, perhaps with a broader or reimagined view of what could be achieved and how.”
The Dædalus issue explores these questions across the life and physical sciences, the science of AI itself, and the broader social, institutional, and epistemological implications of AI-assisted discovery. Among the most interesting contributions are two extended conversations between Manyika and two pioneers of modern AI: Demis Hassabis and Yann LeCun.
Their views suggest two different but complementary ways in which AI could transform scientific discovery. At the heart of their discussion is a fundamental question: Will AI primarily help scientists answer difficult questions, or might it eventually become capable of asking important new questions of its own?
Can AI Ask the Right Scientific Questions?
“Why does science matter to you?” Manyika asked Hassabis.
“I’ve always been driven by a deep desire to better understand the world around us,” Hassabis replied. “The scientific method is arguably the most important idea humanity has ever had to make progress and advance knowledge. For me, there’s nothing more thrilling than being at the absolute frontier of what’s known, pushing the boundaries forward and discovering something new for the very first time — that’s science at its very best.”
But he also emphasized his practical and engineering instincts.
“Instead of getting lost in a theoretical argument, I lean into my pragmatic side and start thinking about how to build tools or experiments to prove hypotheses. AlphaGo and AlphaFold are good examples of that.”
AlphaFold is perhaps the most celebrated example to date of AI’s potential to advance science. By predicting the three-dimensional structures of proteins from their amino-acid sequences, it helped solve a scientific problem researchers had struggled with for decades.
But when Manyika asked Hassabis what capabilities today’s AI systems still lack, his answer focused on something much more difficult: scientific creativity.
“One thing that’s missing and one of the reasons why I don’t think today’s systems are close to AGI yet is that they don’t have true creativity,” Hassabis said. “What I mean by that is, can you have an AI system that doesn’t just solve a difficult problem, which is obviously quite useful and impressive, but can actually come up with a beautiful or meaningful conjecture like the Millennium Prize problems or the Riemann hypothesis? I think that’s much more difficult.”
“Asking the right question is the hardest part of science,” Hassabis added. “What’s the right hypothesis that’s worth exploring and worth doing science on? We don’t fully understand how that level of creativity works in human experts, and we certainly don’t have it yet in our AI systems.”
Today’s most advanced AI systems are increasingly capable of solving difficult problems once those problems have been clearly defined. But the history of science has often been shaped by people who saw a problem that others didn’t see or asked a question that no one had previously thought to ask.
Can AI eventually develop this kind of scientific creativity? It remains an open question, said Hassabis. Meanwhile, he sees enormous opportunities for AI to improve human health, cure diseases, and address energy and climate challenges by developing new materials and technologies, optimizing existing infrastructure, and helping us better understand complex ecosystems.
Looking ahead a few decades, Hassabis offered a highly optimistic vision. “Hopefully the field will have safely stewarded AGI into the world for the benefit of all humanity, and we will be discussing how to navigate a post-AGI world of radical abundance and living in a new golden era of scientific discovery and wonder!”
AI as a Staff of Research Assistants
Yann LeCun offers a somewhat different perspective on the future of AI and science.
Looking back at the extraordinary progress of AI over the past 15 years, LeCun recalled the period when Geoffrey Hinton, Yoshua Bengio and he were trying to rekindle interest in neural networks, which they rebranded as deep learning.
They were surprised by how rapidly the technology was adopted. “Within eighteen months of the first papers showing that deep learning could improve speech recognition, basically every mobile speech recognition system was using a neural network,” LeCun said. “This was my first exposure to how quickly technology is widely adopted when it actually brings something to the table.”
LeCun sees particularly promising opportunities for AI in materials science, chemistry, biochemistry, and medicine. Could AI help us understand and cure chronic diseases and cancers? Could we move beyond predicting protein structures to designing proteins with specific properties? Could AI help create new materials for batteries, energy storage, and catalysis?
But LeCun’s most interesting argument concerns how AI might help scientists understand extraordinarily complex systems.
Traditional science has often relied on reductionism — breaking complex phenomena into smaller components that can be studied and understood individually. But many biological, physical, and environmental systems may simply be too complex to model this way.
“We can use machine learning to train predictive systems that can be phenomenological models,” e.g., scientific models derived from empirical experiments or observations, but not necessarily supported by theory or first principles. “They’re not necessarily models that explain anything, but they can make predictions that are sufficiently accurate and computationally much faster than a traditional reductionist model.”
LeCun is devoting much of his own research to developing AI systems that can find abstract representations of complex phenomena and make predictions within those abstract spaces. “What we do as scientists is to find a representation of the input that eliminates all the details we cannot predict, and we make predictions in that representation space.”
This raises a fascinating question about the nature of scientific understanding. Is it necessary that scientists fully understand the underlying mechanisms of a phenomenon before making useful predictions about it? Or can AI discover abstractions and representations that allow us to predict the behavior of complex systems that no individual human could fully comprehend?
Lecun was then asked whether AI might eventually conduct science and make discoveries on its own, or will it always be a tool for scientists to use for discovery?
“I think it’s going to be a tool for a long time,” he replied. “AI will essentially operate as a staff of diligent research assistants. Even if these AI systems are better than us at solving individual problems or practicing scientific research — and it’s great to work with people who are smarter than you! — we will still be their boss. We will still tell them what to work on, because ultimately, we use these tools to help ourselves as a species or society. Directing what AI systems produce is a role we’ll preserve.”
But both see AI as potentially transforming the scale and speed of scientific discovery. While Hassabis is focused on whether AI can eventually acquire the creativity to ask fundamentally new scientific questions, LeCun is focused on building AI systems capable of finding abstraction representations and making predictions about phenomena too complex for humans to model directly.
An Invention of Methods of Invention
I’ve long been interested in general-purpose technologies — innovations such as the steam engine, electricity, computers, and the internet whose impact extends across much of the economy and society. Over time, AI will clearly join this distinguished group. But Manyika makes a more intriguing argument in his concluding remarks. AI may also be an “invention of methods of invention”— a technology that doesn’t simply improve existing products and processes but helps us discover entirely new ones.
“The interaction between AI and science has two characteristics that together could further expand the possibilities for progress,” he wrote. “AI is among a few technologies with a strong claim to being both a general-purpose technology and an invention of methods of invention.”
What lies ahead might thus be thought of as a combination of the Industrial Revolution and The Enlightenment. The general-purpose technologies of the Industrial Revolution enabled extraordinary economic progress. The Enlightenment and the scientific revolution transformed humanity’s methods for creating knowledge through the scientific method and new instruments of discovery.
AI could potentially combine elements of both. If AI is a general-purpose technology, it could transform industries and economic activity across much of society. But if AI also becomes an invention of methods of invention, its greatest long-term impact may come from accelerating the process by which humanity creates new scientific knowledge and technologies.
Realizing AI’s scientific potential will require developing systems that are more capable of advancing science and making those capabilities available to scientists in many more institutions and countries. At the same time, we will have to address the ways AI could reconfigure the social, institutional, and epistemological foundations of science, as well as the safety and ethical risks that have always accompanied powerful technologies.
The 33 scientists who contributed to this issue of Dædalus don’t offer a single answer to what scientific discovery will look like in the age of AI.
That may be the most appropriate conclusion at this early stage. We don’t yet know whether AI will eventually become capable of making profound scientific discoveries largely on its own. Nor do we know whether machines will develop the kind of creativity required to formulate a beautiful conjecture or identify a scientific question that no human has previously thought to ask.
But AI is already changing what scientists can observe, predict, and explore.
The telescope and microscope didn’t simply help scientists answer existing questions. By revealing worlds that humans had never before been able to see, they enabled entirely new questions to be asked. Modern computing similarly allowed scientists to explore systems whose complexity had previously placed them beyond the reach of human analysis.
AI may prove to be the most powerful instrument of discovery yet. Its greatest scientific contribution may ultimately be not simply helping us find better answers, but enabling scientists — and perhaps someday AI systems themselves — to ask questions we have not yet imagined.