AI & Science: What Is the Future of Discovery?

This essay explores the impact of information processing on chemistry and materials science through a chronological journey that begins with the origins of life and ends with speculations of what will be possible in 2050. We pause at key moments in this history to reflect on changes to the three main, intertwined threads of this story: matter, minds, and tools. “Matter” refers to the chemical environment that our species interacts with. “Minds” refer to the collective of natural and artificial intelligence that we have access to, as well as our inherent biological information-processing capacity. “Tools” refer to the different processing technologies, such as flasks, oil refineries, computer processors, and so on, that humans employ to understand and manipulate matter. Through this framework, the essay considers the roles of cyborgs and androids today and speculates on the potential future of AI for science.

Throughout history, scientific progress has been defined by the tools we build to extend our perception. Today, artificial intelligence stands as our newest instrument, not merely as a computational engine but as a revolutionary lens for deciphering the nature of reality. Drawing on my research journey at Google DeepMind, this essay explores how scientists are deploying AI to overcome the hardest “root node” bottlenecks in science. By applying machine intuition, we are uncovering breakthroughs across biology, materials science, and mathematics, from solving the grand challenge of protein folding with AlphaFold to charting new universes of stable materials with GNoME. The essay concludes by reflecting on the transition toward open-ended, agentic AI systems that actively generate novel hypotheses and their implications for the nature of scientific inquiry. We are evolving from the solvers of intricate puzzles into the architects of profound questions, embarking on the most exciting journey of discovery ever undertaken.

A new paradigm is emerging at the intersection of artificial intelligence and experimental biology, where cells are no longer merely observed, but comprehensively modeled, queried, and predicted in silico. New measurement technologies and the ability to genetically manipulate cells precisely have opened the way to measure cells in extraordinary detail under tens of thousands of perturbations. Concomitantly, AI foundation models are learning to represent, simulate, and even anticipate cellular behavior. This essay traces the convergence of these revolutions, showing how they are giving rise to “virtual cells”: integrative models that unify diverse molecular and spatial data into coherent, functional representations that can generalize across biological contexts and conditions. Beyond representing and interpreting biological lab measurements, virtual cells aim to predict unseen outcomes, imagine new contexts, and guide discovery. In closing the loop between data generation and hypothesis testing, AI is transforming biology into a self-refining, interactive science, resulting in a profound shift: from observing life to actively modeling it, with implications for precision medicine, biotechnology, and the scientific method.

Inspired by the visionary predictions of Jules Verne, this essay proposes that integrating artificial intelligence with a paradigm called “nodal biology” will revolutionize the discovery of treatments and cures. Current drug development is slow, costly, and failure-prone, constrained by the challenge of drug target identification. Solving the “cell perturbation prediction problem”—predicting how human cells respond to any disease-causing perturbation—is key to accelerating successful drug target identification. Nodal biology, the discovery of shared druggable mechanisms (nodes) among seemingly disparate diseases, offers a scalable approach to generate the high-quality data needed to train cell prediction AI models. As an example, a cargo receptor node was identified, linking dozens of genetic diseases and leading to a new drug candidate. The synergistic combination of human scientific intuition and AI-empowered nodal biology is essential for building the biomedical innovation engine of the future, ultimately accelerating treatments for all human diseases.

Enabled by advances in imaging and computer science, the past two decades have featured dramatic progress in the study of comprehensive synaptic-resolution maps of the nervous system. In this essay, we review these achievements and innovations but also argue that subsequent progress will be substantially driven by artificial intelligence. Specifically, we propose that AI is poised to revolutionize the study of brain wiring by: 1) helping to forge a new, functionally grounded definition of “understanding” the brain; 2) enabling sophisticated simulations and predictive models of neural circuits; and 3) identifying subtle connectomic “fingerprints” of neurological and psychiatric diseases. We envision a future in which the joint contributions of connectomics and AI not only decipher the brain’s subtle wiring but also unlock novel avenues for fundamental discovery and therapeutic intervention. 

To ensure a sustainable future, we need to understand how Earth’s climate has changed over time, how different factors have contributed to those changes, and how human action will impact the climate in the future. Developing this understanding involves a continual process of model refinement, calibration, validation, and evaluation against available observations. These tasks present the core opportunity and challenge of applying artificial intelligence to climate: AI is enabling a revolution in the ability to represent the functioning of complex systems, but while the sheer volume of data available to help us understand the earth system is growing at an unprecedented pace, this observational record is often ill-suited to provide robust benchmarks for AI-driven models. In this essay, we present examples that illustrate this tension, focusing on the AI tractability of different applications, with tractability linked to the availability of metrics and benchmarks to guide model development. We also describe a future when developments in AI will accelerate improvement in models used to support climate action and resilience, enabling us to tackle the currently “intractable” frontier.

Artificial intelligence is rapidly being integrated into a variety of tasks in astronomical research, from model fitting to anomaly detection. While AI offers huge potential to accelerate discoveries about the universe, its current capabilities and trustworthiness fall short. I discuss the road to full AI integration into astronomical workflows: why the field of astronomy is a particularly fertile ground for AI development, the potential revolutionary gains possible for scientific analysis and discovery, how AI developments within astronomy can accelerate AI advancements more broadly, and the current roadblocks to fully AI-integrated astronomy. 

Since their origins in the 1950s, cognitive science and artificial intelligence have made slow but steady progress together toward a functional understanding of human intelligence. The arrival of large language models (LLMs) has upended this dynamic, with unprecedented commercial investment driven by the bet that superhuman AI could emerge simply from learning patterns in language at sufficient scale. While language is a singular tool for human thinking, there is far more to intelligence than language, as evidenced by how young children and nonhuman animals learn quickly and robustly even without language, and by the continuing jagged frontier of successes and failures in LLM-based AI. This essay considers another route to intelligent machines, grounded in principled theories and well-tested models of how minds and brains think before language and how learning language transforms thinking. By deploying AI breakthroughs in LLMs as models of language use—rather than as end-to-end models of intelligence—and connecting them to models of the thinking and learning our minds do prior to and independent of language, we have the opportunity not only to build more robust and efficient AI systems, but to rebuild the bridge between cognitive science and AI. This is a route to answering the biggest open questions about how human minds work, and with that understanding, making AI that positively impacts mental health, education, and society in ways unlikely to come from machine learning alone.

Recent breakthroughs in artificial intelligence have been driven primarily by data-centric approaches, including deep learning and large language models. However, despite their remarkable success, relying solely on purely data-driven methods has intrinsic limitations for scientific discovery. In this essay, I argue for knowledge-­centric AI, reasoning grounded in scientific principles, domain structure, and constraints while learning from data. In our research at Cornell University and through interdisciplinary collaborations with researchers across institutions, this agenda has been shaped by data-limited scientific challenges, which has led us to develop knowledge-centric approaches, such as deep reasoning networks, which build interpretable, domain-aligned latent spaces and enforce differentiable constraints end-to-end. Knowledge-centric AI has enabled scientific advances, from automated crystal-structure phase mapping that uncovered high-performing alloyed mixtures to joint species–distribution models that inform conservation, while also delivering general AI methodologies. By aligning with the scientific method and integrating reasoning, learning, and principled experimentation, knowledge-centric AI can catalyze scientific innovation. In our work, this perspective is grounded in computational challenges central to a sustainable future.

Artificial intelligence has made remarkable strides in natural language processing and image recognition, yet its impact on the natural sciences is fragmented. While specialized models like AlphaFold have revolutionized biology, the scientific enterprise remains siloed, with most foundational models narrowly tailored to specific domains or modalities. In this essay, I advocate for a new class of scientific AI: the polymathic foundation model. Inspired by the intellectual versatility of human polymaths, such a model would integrate diverse data types and disciplinary knowledge across the natural sciences. I argue that building such a model is not only technically feasible but epistemologically necessary. I draw on lessons from existing interdisciplinary successes and outline key challenges: scientific dataset curation, multimodal and multitask learning, verifiable knowledge exchange, and interpretability. I close the essay with a cautiously optimistic roadmap for how such models could transform scientific discovery in the next decade.

Scientific progress is limited not by a lack of new ideas but by the time and cost involved in physical experimentation. Scientific discovery is a needle in the haystack problem: it does not help if AI gives you a vastly bigger haystack. Without knowing if any of the ideas work, an AI system that designs experiments just increases the effort required, since performing the experiments to validate the ideas is the real bottleneck. In my view, AI’s most transformative impact in enabling scientific discoveries lies in reducing the need for such experiments. To get there, we need to build AI models that are able to granularly simulate and understand physics at all scales, rather than just abstractly reason in the textual domain. In this essay, I explore what methods like Neural Operators have already helped achieve, what still needs to be done, and what lies ahead.

Artificial intelligence stands poised to transform our society, yet we hardly understand how it works. A synthesis of physics, neuroscience, and AI can fulfill an urgent need: to build a new, unified science of intelligence that explains and improves how intelligence emerges across both artificial and biological neural networks. I discuss four ways this synthesis has begun to and will continue to unfold. First, powerful analytic tools from the physics of complex systems will provide insight into how large neural networks learn and compute. Second, neuroscience will provide clues into bridging the many orders of magnitude advantages that biological intelligence retains over AI. Third, we can go beyond evolution to instantiate neural algorithms in quantum hardware, leading to new devices through AI-led codesign of physics and computation. Fourth, we can meld minds and machines by building digital twins of the brain, yielding insights into not only intelligence but also consciousness and the sense of self through causal modeling and control. Overall, AI will shed light on the nature of our physical and mental realities, raising profound questions about the role of human understanding in the age of AI. 

Advanced AIs of the future will harness the most powerful computational operations known today: quantum computations. The synthesis of quantum computing and AI will give rise to Quantum AI—artificial intelligence with access to quantum computational resources—which is poised to expand the capabilities of intelligent systems. This advancement will lead to transformative abilities, such as ultra­precise sensing, the realistic simulation of complex natural phenomena, the efficient solution of classically intractable mathematical problems, or protocols for secure communication and enhanced coordination in multi-agent systems. The impact of this synthesis will be profound, promising to accelerate scientific discovery in areas from molecular biology to quantum gravity, while empowering engineers to create novel pharmaceuticals, viable fusion reactors, or designer materials with desirable properties. Quantum AI might also help to address a critical bottleneck in classical machine learning: its escalating demand for data and computational power. Beyond such applications lies another even more consequential possibility. If, as some recent theories and data from quantum neurobiology suggest, quantum processes give rise to consciousness and free will, then only Quantum AI—unconstrained by the limitations of classical computation—could enable the expansion of human consciousness and a transcendence of our biological origins.

AI’s most significant contribution to medicine to date is its ability to produce more accurate and comprehensive interpretations of medical images, as validated through randomized clinical trials. Next up is the opportunity to address the steady erosion of the patient-doctor relationship over several decades and a global burnout crisis among clinicians. By reducing the data clerical work of clinicians and giving patients more autonomy, AI has the potential to restore the humanity in medicine and care. Freeing physicians from constant screen time could bring back the physician’s presence during clinic visits and help foster empathy and trust. Furthermore, AI’s ability to aggregate and contextualize a “full stack” of patient data can enable new opportunities for prevention of major age-related diseases, significantly extending human health span. Despite this vast potential, the implementation of medical AI faces substantial challenges, including mitigating bias, ensuring data privacy and security, and establishing appropriate regulatory and reimbursement frameworks. 

The proliferation of artificial intelligence presents Africa’s drug discovery landscape with a transformative opportunity to accelerate research and development (R&D) to address the continent’s heavy disease burden. The integration of AI in drug discovery processes will help researchers by streamlining research pipelines and improving the design of clinical trials so that historically underserved populations can access cost-effective medication tailored to their needs. While challenges include AI skepticism, scarcity of data, and concerns about job losses, sustained advocacy and capacity-building can help to allay these concerns and facilitate widespread buy-in and adoption of AI tools across the continent. At a time of diminished funding for R&D globally, AI-assisted drug discovery—with its potential to alter fundamentally the pace and cost of R&D—should help Africa move toward a sustainable and innovative pharmaceutical paradigm led by Africans, out of Africa, for the global community.

Physics is a sprawling and ambitious enterprise unified by the goal of explaining the natural world, yet it is extraordinarily diverse in how that goal is pursued. The opportunities to harness artificial intelligence to aid in our pursuits to understand the universe are vast. But the heterogeneity of physics in practice is a vivid demonstration of how scientific inquiry varies across scale, infrastructure, and standards of evidence, and how machine learning integrates unevenly as a result. While current applications of AI often focus on accelerating existing workflows, the deeper promise lies in developing robust methods that plug into physicists’ established workflows and toolkits; lower barriers in the continuous cycle between measurement, simulation, analysis, and theory; and create fundamentally new ways of probing, modeling, and controlling physical systems. 

We are in a new era of advancing science through AI. But the tremendous progress is not just about technology: as scientists have adopted AI-based technologies, they have changed how they do science, redefining the role of human intelligence and the human scientist in the scientific endeavor. This essay reflects on some of these changes and some of the current challenges and tensions that science and scientific communities are facing as we grapple with how best to work with AI to advance science and ensure society’s continued trust in scientists and AI-based scientific evidence. 

Uses of advanced artificial intelligence are changing how societies organize labor, govern, produce knowledge, and make meaning. In light of these developments, this essay argues that AI models, tools, and systems pose three interrelated imperatives for social science: they demand renewed attention to social theories of how technology, human experience, and social order are entangled; they require study as objects of inquiry in their own right; and they offer capabilities that may transform—or upend—the practice of social investigation itself. From Weber’s analysis of rationalization to Du Bois’s study of technology and inequality to contemporary scholarship on algorithmic governance, the essay examines what social science distinctively offers: the capacity to historicize the apparently unprecedented, to trace connections across scales, and to center those most affected by technological change. It identifies how algorithmic systems are remaking the distribution of opportunity and risk as a central task of social inquiry and asks what futures social science might help bring into being.

Artificial intelligence is beginning to make remarkable contributions in science, from protein design to materials discovery and experiment design. Yet we believe that moving from powerful tools to autonomous scientists is not just a technical challenge but also a deep philosophical and conceptual one. For that reason, we argue for a Philosophy of Autonomous Science (PAS): a program that translates core epistemic aims, including understanding, curiosity, surprise, interest, creativity, and novelty, into computable, nonanthropocentric objectives, while ensuring the safe and successful deployment of artificial scientists. Drawing on our own research experience, we show how insights from the philosophy of science can directly inform the design of AI systems for scientific discovery; outline ten inaugural questions that define the agenda of PAS; and invite philosophers, scientists, and AI researchers to collaborate in shaping the principles of the coming era of artificial scientists.