Which Nobel Prize Questions AI Actually Made Tractable

The Nobel Prizes reward the kind of discoveries that, in retrospect, seem almost inevitable — the right insight arriving at the right moment, occasionally fifty years before the tools existed to confirm it. By 2029, several questions that looked like they might occupy scientists for another generation have been cracked open, or at least meaningfully dented, by AI systems. The pattern of which questions yielded is not random, and understanding that pattern is more useful than any specific breakthrough.

There’s a class of scientific problem that is hard primarily because it requires searching an enormous space of possibilities with some evaluable criterion for success. Protein folding is the archetype: the space of possible three-dimensional conformations is astronomical, but the criterion (lowest energy state, consistent with physical chemistry) is well-defined and can be computed. AlphaFold’s success rested on this structure. The problem was hard for humans not because the underlying physics was mysterious but because the search was intractable without the right inductive biases, and deep learning learned those biases from the existing structure database.

Materials science has followed a similar pattern. The discovery of new superconductors has been agonizingly slow for decades — each candidate required synthesis and testing, and the theoretical guidance for which crystal structures might superconduct at higher temperatures was limited. AI models trained on known superconductors and their structures have now proposed, and experimentalists have confirmed, three new families of materials that remain superconducting at temperatures above what was previously thought possible without exotic conditions. The 2028 Physics Nobel recognized this work. Importantly, the AI didn’t derive the physics from first principles; it identified patterns in a large empirical dataset that human scientists hadn’t recognized. That’s a specific kind of scientific contribution, and it’s worth being precise about what it is.

The Nobel Prizes that AI has not yet cracked are equally revealing. The problem of consciousness — why physical processes in neural tissue produce subjective experience — remains untouched. This is not surprising. The problem is hard because there’s no agreed criterion for success. You cannot train a loss function on “is this an explanation of subjective experience” because no one agrees what a satisfying answer looks like. AI is very good at optimizing against well-defined objectives; it has no purchase on questions where the objective itself is contested. The same applies to interpretations of quantum mechanics. The calculations are not the question. The question is what those calculations mean, and that’s philosophy, not search.

Neuroscience sits in an interesting middle position. Connectomics — the mapping of neural circuits — has been dramatically accelerated by AI image analysis applied to electron microscopy data. The full wiring diagram of the mouse visual cortex was completed in 2026, a project that would have taken multiple human lifetimes without automated segmentation and synapse identification. But having the wiring diagram and understanding what it does are different things. The wiring diagram is data; the explanation of visual processing is a theory. AI provided the data. The theories are still being argued over in competing papers.

Drug-receptor binding affinity prediction is another domain where AI has moved a long-standing Nobel-adjacent question from “approximately intractable” to “tractable with caveats.” For decades, computational chemists could roughly rank small molecules by predicted binding affinity to a target protein, but the error bars were large enough to make prioritization unreliable. Current AI-based methods, trained on the enormous datasets generated by high-throughput screening programs, have substantially reduced those error bars for well-characterized protein families. The caveat is that performance drops significantly for proteins outside the training distribution — precisely the novel targets where good predictions would be most valuable.

Climate science presents a different case. The fundamental physics of climate has been understood for over a century. The limitation on useful climate prediction has been computational: the atmosphere and ocean interact across spatial scales from millimeters to thousands of kilometers, and no model can resolve all of them simultaneously. AI-based parameterizations — learned approximations of small-scale processes that are too fine-grained to simulate explicitly — have improved the fidelity of climate models in ways that matter for regional prediction. This is not a Nobel Prize-level breakthrough in the sense of a conceptual advance; it’s an engineering advance that has made existing theory more useful. Climatologists are generally careful to maintain this distinction even as they celebrate the practical improvement.

The replication crisis, which predated the AI era but has intensified with it (more on that elsewhere), has made the scientific community appropriately cautious about pattern-recognition results that lack mechanistic explanation. An AI finding a correlation in a large dataset is not the same as understanding why the correlation exists. The best researchers have internalized this and use AI findings as hypotheses to be explained mechanistically, not as conclusions. The Nobel committees have, so far, rewarded this framing: the prizes have gone to teams that used AI to find something genuinely new and then did the hard work of understanding why it was true.

The question of which Nobel Prize questions AI will crack in the next decade is worth thinking about carefully. The candidates are the domains where large amounts of well-structured data exist, where success criteria are evaluable, and where the limiting factor is search efficiency rather than conceptual understanding. Genomics has enormous amounts of well-structured data; the mapping of regulatory sequences and their effects on gene expression is a domain where AI will likely produce several more important findings. Astrophysics, which generates data at a rate that exceeds human analysis capacity, is another candidate — pattern recognition in gravitational wave data, in transient survey data, in spectra from the next generation of telescopes.

What AI has done, and will continue to do, is remove certain categories of practical difficulty from science. The difficulty of handling massive datasets, of searching large combinatorial spaces, of recognizing patterns in high-dimensional data — these are all diminishing as barriers. What remains, and what will remain, is the harder intellectual work: formulating the right question, deciding what evidence would constitute an answer, and explaining why the answer is what it is. The prizes that come from AI will be for the scientists who understood how to ask AI the right question. That’s a real scientific contribution. It’s just not the kind that machines can do on their own.