The video is just over an hour long. You can watch it here.
It was filmed on the 26th of May this year at the Royal Society in London, part of the Nobel Prize Dialogue series — an event the Nobel Foundation runs annually to put prize winners into conversation with each other and with researchers working on the sharpest edges of adjacent fields. This particular conversation brought together three people: Demis Hassabis, who won the 2024 Nobel Prize in Chemistry for AlphaFold and is CEO of Google DeepMind; Paul Nurse, who won the 2001 Nobel Prize in Physiology or Medicine for his work on cell cycle regulation and is President of the Royal Society; and Alison Noble, Technikos Professor of Biomedical Engineering at Oxford and chair of the Royal Society's Science in the Age of AI working group. Adam Smith from Nobel Prize Outreach moderates.
I've watched a lot of panel discussions about AI and science. Most of them disappoint, for a predictable reason: the participants haven't actually done science at the frontier and don't have a technical command of what these tools actually do. What you get is one side arguing AI will solve everything, and the other side arguing we need to be careful, and neither side has enough depth to get past the slogans. This discussion is different. These three people have done the work. They have the right kind of calluses. And they are willing to sit in the discomfort of genuinely not knowing the answer to the question they're being asked. That combination is rare enough to be worth your time.
The question that organises everything else
The title — "The Future of Science With AI" — is generic in the way event titles have to be. But the question the panel actually circles is more specific and harder than it sounds.
It is not: Can AI do science? That question is substantially answered. AlphaFold solved a fifty-year-old grand challenge in structural biology — predicting the three-dimensional structures of proteins from amino acid sequences — and made over 200 million such predictions freely available to more than three million researchers in 190 countries. GraphCast outperforms traditional numerical weather prediction at ten-day lead times. DeepMind's Gemini Deep Think model reached gold-medal performance at the International Mathematical Olympiad in 2025. The answer to "can AI do science" is: it already is, and in some narrow domains at a level that is beginning to exceed what the best human specialists can consistently achieve.
The question the panel keeps returning to is sharper: what is the role of human scientists once AI can do a significant fraction of what scientists currently do? And inside that, a second question that I think is the most important one in the room: is there an irreducible core of scientific inquiry that AI cannot perform, and if so, what is it?
Hassabis has a clear answer to the second question, one he's been consistent about across multiple contexts. He laid it out in detail in a Daedalus interview published by the American Academy of Arts and Sciences just before this panel, and it surfaces throughout the discussion at the Royal Society: the irreducible core is asking the right question. "Conjectures and hypotheses are going to come from the human experts," he says. The hardest part of science is not solving the problem. It is knowing which problem is worth solving — identifying the next meaningful hypothesis, recognising the question that, if answered, opens an entire new branch of research rather than merely extending a dead one.
AI systems today, including the most powerful ones, can be extraordinarily capable within a well-defined problem space. What they are not yet capable of is the kind of intuitive leap that defines a Ramanujan or an Einstein — the ability to propose something simultaneously new and meaningful without being given the question first. Hassabis describes this as the difference between solving a difficult problem (impressive, increasingly within AI reach) and formulating one worth solving (the thing he believes defines scientific genius and which current systems don't have). He offers a specific test: take an AI system with a knowledge cutoff of 1910 and see whether it arrives at General Relativity, as Einstein did, by 1915. The answer today is clearly no. That is not a small gap.
This matters more than it initially appears, because it reframes both what we should worry about and what we shouldn't. The anxiety about AI "replacing" scientists tends to focus on the wrong stage of the process. Yes, AI will automate a great deal of what researchers currently do: literature synthesis, data cleaning, hypothesis testing within defined parameter spaces, routine analysis pipelines, low-level code, experimental design optimisation. It is already doing some of these things well, and it will do all of them better than any individual researcher working unaided within the next five years. But that, in Hassabis's framing, is not the essential work. The essential work is knowing what question to ask, and deciding what an answer would actually mean. That is what the best scientists spend most of their time doing, and it is the part that remains, for now, beyond what we know how to give to a machine.
What Alison Noble brings that the panel needs
Alison Noble is important here in a way the panel structure does not fully foreground. She's not merely a biomedical engineer who works with machine learning — she's the person who chaired the working group that produced Science in the Age of AI, the Royal Society's 2024 report on what AI is actually doing to scientific practice. That report is, in my view, the most careful evidence-based analysis produced by any major scientific institution on this subject, and Noble brings its findings into the room as a useful corrective to the enthusiasm that tends to dominate these conversations.
The report's central finding is nuanced in a way that popular coverage rarely captures. It does not say AI is bad for science. What it says is that the widespread adoption of opaque, poorly-understood, often proprietary AI tools into scientific pipelines is introducing new vectors for irreproducibility, and that the scientific community has not yet built the technical, cultural, or regulatory infrastructure to catch those failures.
This is a serious and immediate problem, and it gets too little airtime in the standard AI-in-science conversation. When you use a model you don't understand, trained on data you can't inspect, to generate a finding that a colleague needs to reproduce using a commercial tool they may not have access to — you have not broken the scientific method, but you have undermined several of the conditions that make it function. The finding may be real. It may also be an artefact of the model's training distribution, or of preprocessing choices you made implicitly, or of a dataset that is no longer available because the commercial provider updated it quietly. The Royal Society report found a growing body of AI- and ML-based studies whose results could not be independently reproduced — not because the science was dishonest, but because the tools were opaque and the data was proprietary and the workflow was not documented at a level that would allow a third party to reconstruct it.
Noble's position in the panel, then, is not that AI is dangerous. She is a machine learning researcher herself. Her position is: we need to build the scaffolding of good practice around these tools at the same speed as the tools are being deployed, and right now we are not doing that. There is a lag between adoption and rigour, and that lag is creating a reproducibility debt that science will have to pay off eventually, in the form of retracted papers, missed findings, and eroded public trust. Better to address it now.
The report made several concrete recommendations that are worth naming:
Open science frameworks for AI-based research. If your finding depends on an AI model, that model and its training data should be accessible to others attempting to reproduce it — or, where that's not commercially possible, at minimum there should be a documented, tested alternative pathway. The current situation, in which findings routinely depend on models and datasets that are neither open nor documented to the standard required for reproducibility, is one we would not tolerate in any other methodological domain. We require authors to disclose their statistical software and its version. We require them to share their raw data in most journals. The same discipline needs to apply to AI components of a scientific workflow.
CERN-style shared computing infrastructure. One of the starkest inequalities in contemporary science is compute access. A team at a well-funded institution with a commercial AI partnership can run analyses that a team at a lower-resourced one simply cannot afford to replicate. The report called for investment in regional and cross-sector compute infrastructure — shared, like CERN for particle physics — to prevent AI from deepening existing resource inequalities in science. I feel this acutely in my own work in low-resource clinical settings: the gap between what I can do with the tools I have access to and what a team at a major AI laboratory can do is not primarily a skills gap. It is a capital gap. And it is widening.
AI literacy, not just AI skill. There is a meaningful difference between knowing how to run a model and knowing enough about how it works to interpret its outputs critically. Many researchers using AI tools have the former without the latter, which is dangerous precisely because capable-looking outputs from opaque models can be confidently wrong in ways that are hard to detect without deeper understanding of the architecture and its failure modes. Building that literacy at scale is a training and culture problem, not a technology problem, and it requires investment in education that current policy timelines are not yet matching.
Paul Nurse and the longer view
Paul Nurse's contribution to this conversation is different in character from Hassabis's or Noble's, and it anchors the discussion in a way I found valuable. He is a classical biologist who did bench science — who spent years working out how the cell cycle is regulated, what the molecular switches are, how a cell knows when to divide. That work earned a Nobel Prize and it came from the slow, careful, hypothesis-driven mode of science that most AI-in-science conversations quietly push aside as they describe the automated future.
What Nurse keeps returning to, as I read it, is what happens to the epistemic structure of science — not just its productivity — when a layer of learned abstraction is inserted between experiment and understanding.
Science at its best is not simply a mechanism for producing correct answers. It is a mechanism for producing correct answers with explainable reasons, such that you can derive new predictions from those reasons and test them elsewhere. A model that correctly predicts protein structures but whose mechanism of prediction is fundamentally opaque does not give you understanding in that deeper sense. It gives you an extraordinarily powerful lookup table. That is genuinely useful — transformatively so, as AlphaFold has demonstrated — but it is a different kind of knowledge, and conflating it with mechanistic understanding can mislead you when you try to extrapolate to a situation the model wasn't trained on.
This concern connects directly to what Noble is saying about reproducibility. When you don't know why your model gives the answer it does, you don't know when it will be wrong. In science, knowing when your method fails is as important as knowing when it succeeds. A method with unknown failure modes is a method you can't fully trust, even when it works.
Nurse's implicit question to the panel — one I don't think gets a fully satisfying answer in the hour they have — is whether the increasing use of AI in science is shifting our epistemic standard from "do we understand this" to "does this work," and whether that shift, which is pragmatically seductive in the short term, carries costs for the long-term integrity and productivity of scientific knowledge.
The proprietary problem
One thread in the panel discussion that I haven't seen picked up much in coverage of the event is the proprietary nature of the most capable AI tools and what that means for science as a collective enterprise.
The most powerful models for scientific work are not open. They are built and owned by a small number of large commercial AI laboratories, and accessing them means depending on those companies' commercial decisions, pricing structures, model update schedules, and terms of service. This introduces fragilities into the research enterprise that are new and that we have not yet developed good institutional responses to.
First, reproducibility. If your scientific finding depends on a specific version of a large commercial model, and that model is updated, deprecated, or discontinued, the specific capability your work depended on may no longer exist. This is not hypothetical: commercial model versions are routinely retired on timescales of months. Researchers trying to reproduce your work in five years may be unable to, not because you did anything wrong, but because the tool changed underneath you and the old version is no longer available.
Second, access inequality. Commercial AI capabilities follow commercial pricing. The most powerful tools are most accessible to the most well-funded institutions, which means AI is reproducing and potentially amplifying the structural inequalities that already exist in global science. The history of scientific infrastructure — from particle accelerators to genomic sequencing — suggests that the solution to this is not to wait for prices to fall, but to invest in shared infrastructure. We are not yet doing that for frontier AI on anything like the scale required.
Third, and more subtly: agenda-setting. Who decides which scientific problems the most capable AI tools are optimised for? If the primary accountability of the companies building and maintaining those tools is to their shareholders, there is a real question about whose research agenda those tools end up serving. Drug discovery for diseases of the wealthy is profitable. Drug discovery for neglected tropical diseases is not. The tools are not neutral. The scientific community needs to be actively shaping how they are deployed and governed, not just using them as given.
Radical abundance and its discontents
Hassabis has spoken at length in other forums — and this comes through in the panel — about a vision he describes as "radical abundance": a world in which AI has successfully bottled the scientific method and can run the hypothesis-experiment-analysis cycle autonomously, at a pace and scale that makes the scientific productivity of the last century look like a rough draft. He frames this as a 10-to-15 year horizon, leading to what he calls "a new golden era of discovery — a kind of new renaissance."
This is not a fringe view. It is the operating premise of Isomorphic Labs, the DeepMind spinout that Hassabis leads in parallel, which has raised over two billion dollars explicitly to apply the AlphaFold methodology to drug discovery with the stated goal of eventually solving all disease. It is consistent with what Dario Amodei at Anthropic described as a "compressed 21st century" — the idea that AI accelerates biological and medical progress by a factor of perhaps ten to fifty, collapsing into a decade the advances that would otherwise take a century.
But sitting in that vision alongside Nurse and Noble introduces a useful tension. Radical abundance — AI generating scientific insights at extraordinary speed — presupposes that the insights are trustworthy. It presupposes that the underlying data is clean, that the models are interpretable enough to catch their own errors, that the scientific community has the AI literacy to know when to trust what they're being told, and that the incentive structures around publication and novelty don't reward the confident wrong answer over the carefully hedged right one. None of those conditions currently obtain reliably. They are things we need to build.
The pace of discovery that Hassabis envisions is not incompatible with the rigour that Nurse cares about or the infrastructure that Noble is calling for. But they need to develop in parallel, not sequentially. A scientific literature full of AI-generated findings that nobody can reproduce is not a golden era of discovery. It is a different kind of problem, and one that would set the actual golden era back considerably.
Where I land on this
I'll say plainly what I think, in the context of my own work at the intersection of AI and clinical research.
Hassabis is right that the immediate horizon for AI in science is augmentation rather than replacement. The tools available now can make individual researchers faster, can find patterns in datasets too large for any individual to analyse manually, can automate the parts of research that are laborious rather than creative. In clinical AI projects — including the kind of work I've been building with diabetic retinopathy screening — the hard work is still the hard work: deciding what the right question is, designing a study that will give you an answer you can trust, interpreting findings in clinical context, and knowing when your model is behaving badly. None of that has been automated. I don't believe it will be automated on any timescale that should change how I approach my work this year.
But Noble is right that the scaffolding problem is serious and immediate, and it is one I recognise in my own practice. The tools are moving faster than the habits around them, and I use AI in ways that introduce failure modes my existing quality-control habits were not designed to catch. That is not a reason not to use the tools — the gains in efficiency are real and matter enormously in resource-constrained settings. But it is a reason to be more rigorous about methods sections, more honest about model limitations, more invested in open-science practices around the AI components of my workflows, and more willing to distrust a confident-looking output until I have some independent reason to believe it.
And Nurse's point lands with me too, in a personal way. There is a seductive quality to confident outputs from capable AI systems that can short-circuit the slower work of actually understanding why something is true. I've noticed this in myself. The slowness of that understanding is not a bug in the scientific process — it is load-bearing. It is what allows you to extrapolate, to know when to trust and when to doubt, to build on a result in a new direction. The discipline of demanding to know why, not just what, is one I need to apply to my own use of AI tools, not just recommend to others.
The irreducible core
I started by asking whether there is an irreducible core of scientific inquiry that AI cannot yet perform. After watching this panel and thinking about it for longer than it took to watch it, my tentative answer is: the irreducible core is judgment under genuine uncertainty about what matters.
Not uncertainty about a fact you could look up. Not uncertainty about which of two well-defined hypotheses is better supported by the available data. The more primal uncertainty: is this the right question? Is this the right problem? Does this answer, if I get it, actually advance understanding, or does it produce a finding that is locally correct and globally misleading?
That judgment is not mystical. It is built from deep domain knowledge, from a developed aesthetic sense of what an interesting result looks like, from a grasp of the literature that goes beyond what any model trained on that literature can reconstruct — because the trained model inherits the field's consensus, while the best scientists are precisely the ones who know where the consensus is wrong — and from a kind of intellectual responsibility, a sense of being accountable for whether the question matters, that has no clear analogue in a system minimising a loss function.
How long does that irreducibility hold? I don't know. Neither does Hassabis. He has been admirably honest about this: the exact moment at which AI systems develop genuine scientific creativity — the ability to ask the right question rather than just answer it — is not something he claims to be able to predict, and he thinks it may require breakthroughs we haven't had yet. Noble is more focused on the near-term infrastructure gap and less willing to speculate about timelines. Nurse's concern runs deeper, to the question of whether "bottling the scientific method" is even the right thing to aim for, or whether the slowness and humanness of science is not a limitation to be engineered away but a feature that keeps knowledge honest.
What I take from this panel is not an answer but a better-calibrated set of questions — and a clearer sense of what we need to be building alongside the tools themselves. The tools are coming. What shape they land in, what the scientific enterprise looks like on the other side, and whether the knowledge they generate will be trustworthy and equitably accessible — those are not questions the technology answers by itself. They are the questions we are responsible for.
Watch the full panel discussion here: Demis Hassabis, Alison Noble and Paul Nurse Discuss the Future of Science with AI — Nobel Prize Dialogue in partnership with the Royal Society, 26 May 2026.
The Royal Society's Science in the Age of AI report (2024) is available in full at royalsociety.org. The Daedalus interview with Demis Hassabis, "AI as the Ultimate Tool for Science," is open access at amacad.org.