Every few months, a headline says AI has proven a theorem or solved a millennium problem in mathematics. Then I hear that physics is next and quantum gravity is cited as what they want to "solve". That's a statement from someone who doesn't understand science. I read these the way I read headlines about string theory, with one eyebrow raised and a cup of coffee in hand.
AI is going to do a lot for physics. It will not settle a question about nature from inside a computer, and neither will a theorist with a pencil. A theorist with a pencil can write down and work out a new idea. AI can make sure that idea does not violate a proven mathematical result and check for consistency with known data. The universe as observed in the lab and in the heavens gets the last word. That's how science works.
Math is not Physics and Physics is not Math
Has AI proven the Riemann hypothesis? There is a claim today that it has. A valid proof settles the question inside mathematics, whatever the universe happens to do. We can write down spaces with any number of dimensions, geometries with any curvature, and algebraic structures with no known physical counterpart. Mathematics can describe worlds that look nothing like ours.
Physics has to describe the one we live in. A mathematically consistent theory can describe a perfectly coherent universe that is simply not this one. The graveyard of physics is full of beautiful mathematics that did not survive contact with experiment, and more of it sits in limbo with no known physical relevance. For that reason, solving a mathematical problem tied to quantum gravity does not solve quantum gravity as a physical problem.
Physics Is Tested, Not Proven
A physical theory is never proven the way a theorem is proven. It gets tested against observation, and it survives or it fails. A theory can be beautiful and mathematically airtight and still be wrong about the universe we live in.
General relativity shows how this goes. Einstein used it to calculate the anomalous precession of Mercury's orbit, about 43 arcseconds per century. Astronomers had known about that discrepancy for decades, so this was a postdiction. It was a very good one, because the number came out of the theory with nothing adjusted to make it fit. Even so, physicists wanted a real prediction.
General relativity supplied one. Starlight passing near the Sun should bend by a specific amount, twice what a Newtonian estimate gives. In 1919 Arthur Eddington, Frank Dyson, and their colleagues observed a solar eclipse to check. Later observations with far better instruments have confirmed the effect many times over.
That is how a theory earns its standing. It takes years, and it takes more than one group.
Explaining a result we already know and calling a new result in advance carry different weight. A good scientist keeps track of which one they have done.
So any headline saying AI has "solved quantum gravity" deserves heavy skepticism. Humans have already built sophisticated mathematical frameworks for quantum gravity, among them string theory and loop quantum gravity. What we lack is experimental evidence that picks one of them, if any is right. An AI could write down another candidate, and maybe a better one. A candidate that agrees with existing data still has to face data nobody has taken yet. At best, a headline like that means the AI solved some open theoretical problem. At worst it is hype.
A responsible claim sounds different. A new theory has been proposed, it agrees with known physics and observations, and it makes predictions that experiments could test. That is how physics works.
Where I Am Coming From
I am a theoretical physicist who uses AI and has helped train it. Through platforms such as Outlier and Mercor I have done work for top AI labs, including teaching models to reason through problems in mathematics and physics and to judge arguments about proving or disproving statements. These models know the physics they know in part because of people like me.
I am also a member of the LISA Consortium. LISA, the Laser Interferometer Space Antenna, is a gravitational wave observatory being built by the European Space Agency with NASA and other partners. I do NASA Space Grant research, and I teach physics and astronomy as a part-time lecturer at Chicago State University.
I use AI in my own work to check mathematics, write code, and look at data. I have learned from AI, and AI has learned from me. The AI you use do to physics can do it in small part because of learning from me. How many people can say that? When I push back on the hype, I do it as someone who likes the tool and has seen where it stops.
AI Could Run Experiments Someday
Future AI systems could do more than propose theories and crunch data. They could design experiments, run lab equipment, take measurements, and adjust the procedure as results come in. Plenty of labs already let computers control very sophisticated instruments. I see no fundamental reason an AI could not take part in experimental physics.
Even then, somebody has to build the apparatus. Engineers, machinists, technicians, and experimental physicists make the detectors, install the optics, calibrate the sensors, and keep everything running. An AI can reason all day and still get no new evidence without instruments that touch the physical world. Maybe robots will do much of that work too. The requirement stays the same. A physical theory needs a physical test. An AI might run the experiment, and the experiment is what settles the question.
Right now, an AI lab announcing that they have solved a fundamental physics problem is only showing that they fundamentally do not understand physics.
Postdiction, Prediction, and the Training Data Problem
Suppose an AI lab announces that its model produced a new theory which correctly predicts a phenomenon we have already observed. Was that a prediction?
I would want to know whether the observation was in the training data. I would want to know whether the developers used it, directly or indirectly, to evaluate and improve the model. I would want to know how many candidate theories they tried before they published the one that matched. A model trained on a huge share of the scientific literature has seen things its own developers may not be able to list.
Postdiction is a legitimate part of theoretical physics. A theory that cannot explain what we have already measured is in trouble from the start. The trouble with AI postdiction is leakage and tuning. If the answer may have been in the training set, matching it tells us very little.
A stronger demonstration would be a specific prediction made before the data exist, with numerical values and uncertainties, recorded publicly ahead of time and then checked by independent experiments. If an AI does that for a genuinely new phenomenon, I will say so in print. Then we ask the universe.
A Program to Propose and Test Quantum Gravity
Instead of asking AI to solve quantum gravity, I would like to see a research program where AI helps physicists build hypotheses, check them against what we know, and find the experiments that can tell competing theories apart. I will not assume any particular answer to quantum gravity here.
The AI would need comprehensive access to established physics: classical mechanics, electromagnetism, general relativity, quantum mechanics, quantum field theory, the Standard Model, and the mathematics under all of them. It would also need access to the evidence, from accelerator experiments, precision measurements, astrophysical observations, gravitational wave data, and cosmological surveys. Whenever possible it should work from raw data and not only from published summaries. That does not mean stuffing every byte into the training. It could draw on curated databases, likelihoods, simulations, and analysis tools as needed. Machine learning already does part of this in collider analysis and cosmology, mostly for classification and parameter estimation. I am describing an extension of that work.
Physicists could then use the AI as a hypothesis testing machine. A physicist proposes a modification of general relativity, a new quantum field, or a different structure for spacetime. The AI checks it. Is it mathematically consistent? Does it reduce to established physics where that physics has been tested? Does it break conservation laws, introduce unphysical degrees of freedom, or go unstable? Does it conflict with observations? A proposal that fails can be rejected or revised quickly, and that alone would save researchers a great deal of time.
We should not turn established physics into dogma. A new theory may challenge an existing principle. It then has to explain why experiments have supported that principle so well, and it has to offer differences we can test.
The theories that survive would be developed into testable models. The AI could calculate the signals, work out the measurement precision needed, and estimate the resources required. Then the proposed tests could be sorted:
Bench experiments that existing university or industrial labs could run.
Astronomical and astrophysical observations using telescopes, gravitational wave observatories, pulsars, black holes, or cosmological surveys.
Accelerator experiments at existing or reasonably foreseeable facilities.
Combined programs that need several kinds of measurement.
Tests that work in principle but the test are all impractical.
That last category deserves suspicion. If the only distinguishing prediction of a theory needs an accelerator as big around as Earth's orbit, I do not call that theory testable. It is nearly as bad as being untestable in principle. Someone may find a clever indirect test, and gravitational physics already uses astrophysical sources to probe energies no accelerator can reach. Until somebody does, elegance should not be mistaken for confirmation.
A program like this would sort candidate theories by what we can test now, what we might test soon, and what lies out of reach. That would help fundamental physics whether or not the AI ever came up with a new theory of its own.
What Einstein Actually Did
Here is something I see in current AI systems. They are very good at combining existing mathematical ideas. They interpolate between established theories, push equations around, and spot similarities between structures. What I have not seen is a reliable conceptual revolution, the kind that changes what the mathematics means physically.
Look at special relativity. Hendrik Lorentz and others already had mathematical relationships that accounted for many results in electrodynamics, including the failure to detect motion through the ether. Einstein started from two physical postulates: the laws of physics are the same in every inertial frame, and light in a vacuum has the same speed for every inertial observer. From those he derived the relationships between space and time that we call the Lorentz transformations.
The interpretation is what matters. Length contraction and time dilation stopped being patches added to explain a failed ether search. They became consequences of how space and time are built. The same mathematics described a different reality.
The 1900 Challenge
Here is a challenge for any AI lab that thinks its models are close to changing fundamental physics. Build a model that knows nothing about physics discovered after 1900, and see whether it can find special relativity on its own.
Telling a modern chatbot to pretend it is 1905 will not work. Today's models know Einstein's papers, the Lorentz transformations, Minkowski spacetime, and a century of textbook explanations. Asking one of them to recreate special relativity tells us nothing about whether it could have discovered the theory.
The model has to be built for the test. I suggest a cutoff of January 1, 1900. A cutoff of 1905 would put Lorentz's 1904 work and Poincaré's papers inside the training data and spoil the exercise. The model gets the mathematics and experimental results scientists had at the time: Maxwell's equations, classical mechanics, and the Michelson-Morley experiment. Everything appears in period form, with the notation, uncertainties, and assumptions of the day. No modern textbook explanations of relativity. No modern talk of spacetime. No exercises that smuggle in the answer.
Then pose the problem. Explain the Michelson-Morley result and reconcile it with Maxwell's equations and the principles of mechanics. Let the model form hypotheses, find contradictions, and propose experiments that would tell its explanations apart. I would not demand Einstein's wording or his route. Science does not have to repeat itself. I want to see whether the model reaches the physical principles of special relativity, or another framework that predicts just as well, and whether it can say what evidence would separate its theory from the rivals.
If the model arrives at something like Lorentz's ether theory, I would count that as a reasonable result. A good scientist knows when the evidence cannot separate two interpretations.
There is a practical objection. Text from before 1900 is nowhere near enough to train a model the size of today's systems. I think that fact tells us something. Einstein worked from a small body of evidence and a great deal of thinking, while today's models draw on an ocean of text. If the test calls for a smaller model, or a different way of training, then that is what the test calls for, and the result would still be worth having.
The challenge could repeat across the history of physics. Could a model trained on nineteenth century knowledge find the principles behind quantum mechanics? Could it find a working theory of atomic structure? Could it infer a new principle and go beyond fitting known observations with ever more elaborate equations? Success would not show that AI can solve every problem in physics. Failure would not show that it never can. The results would still tell us far more than a demo where the AI reproduces an answer that was already somewhere in its training data.
I have worked with enough models to doubt that today's would pass. I would love to be wrong. If an AI company wants to prove me wrong, I will hand one of its billionaire owners a crisp $20 bill.
The Universe Gets the Final Say
AI has real promise for science. It can help researchers analyze data, explore models, catch inconsistencies, run simulations, and design experiments. It may someday run laboratory equipment. It may even find physical principles no human has imagined.
The standards stay the same whether a theory comes from Einstein, a graduate student, or a model running in a warehouse full of computers. It has to be mathematically coherent. It has to account for what we have observed. It has to predict something. And those predictions have to meet the physical universe. Even a theory that passes every test so far stays open to the next one.
Quantum gravity will not be settled by an elegant set of equations or a good press release. Maybe AI will help us find the theory. Maybe it will help design the experiment, run the equipment, and analyze the results. The final answer will still come from the universe.
Claude pushed back on its limitations.
I worked with Claude Sonnet 5.5 and ChatGPT on this. Claude in particular did not really like this but then I asked it, at the second-highest effort setting, to come up with a new idea. They likely have an internal model that would approach this challenge in a better way but fundamentally would have the same problem. My words are bolded.
I made my statements more definitive. AI as we now speak of in LLM's such as yourself show no proof of being able to come up with new ideas.
To prove me wrong, propose a new set of 1-5 postulates or axioms which would allow constructing a theory of quantum gravity.
Here is my attempt. I think it will mostly support your point, and I’ll say why after the postulates.
Five postulates
- Finite relational states. Every region bounded by a causal horizon has a finite-dimensional Hilbert space, and the log of its dimension is the horizon area divided by four Planck areas. States are defined relative to subsystems, with no global observer.
- Geometry from correlation. There is no fundamental background metric. The distance between two subsystems is a decreasing function of their mutual information.
- Bounded coupling. Each subsystem interacts directly with only a finite number of others, and the dynamics are unitary with a Lieb-Robinson-type speed limit that becomes c in the continuum.
- Horizon thermodynamics. For every local causal horizon, the heat flow across it equals temperature times the change in entropy, using the entropy from postulate 1.
- Nonclassical mediation. The gravitational interaction between two massive systems is carried by degrees of freedom that can become entangled. Gravity cannot act as a purely classical channel on superposed sources.
What the set is meant to do. Postulates 1 and 4 together give the Einstein equations as a long-distance equation of state. Postulates 2 and 3 are meant to produce locality and dimensionality as emergent features. Postulate 5 is the one that can fail in a lab.
Tests, using your categories
- Bench: Postulate 5 predicts gravitationally induced entanglement between two small masses in superposition, the experiments proposed by Bose et al. and by Marletto and Vedral.
- Astrophysical: Postulates 2 and 3 risk violating Lorentz invariance at some scale. Gravitational wave dispersion limits from GW170817 and future LISA data constrain that directly.
- Accelerator: nothing practical.
Why this doesn’t prove you wrong
- Each postulate has a lineage. Postulate 1 comes from Bekenstein, Bousso, Banks, and others. Postulate 2 follows Van Raamsdonk and Swingle. Postulate 3 comes from quantum cellular automata work. Postulate 4 is Jacobson’s 1995 argument, and Postulate 5 is the Bose and Marletto-Vedral proposal. What I did was select and combine published ideas, which is the interpolation you describe. Whether the combination is new, I cannot verify, since I cannot search my own training data.
- I have not derived anything. The hard step is getting Lorentz invariance out of a discrete, bounded-degree structure. Simple versions of that fail against existing dispersion limits, so this is a program and not a theory.
Where I’d push back on the essay. The evidence supports “no demonstrated novel physical idea from an LLM.” It does not support “cannot, even in principle.” Your own 1900 challenge says a failure would not establish impossibility, so I’d match the strength of the headline to what the test can show. That is your call, and it is your essay.
Your 1900 test is also cleaner than this one. Nobody can check my postulates for novelty against a model that has read the whole literature. A model that has not read it removes that doubt.
If you want, I can run these five through your own filter: consistency checks, the Lorentz invariance problem, and what each test would need to distinguish this from the existing approaches.
That and your model s more or less a version of the computational/simulation hypothesis. Where finite volumes of space-time "compute" what is happening based on what is going on in other volumes. A real soltuion would have to be something more fundamental than space time itself.
You’re right about the central flaw, though I’d draw the line slightly differently than “simulation hypothesis.”
The structural problem is that my postulates assume the thing they claim to derive. Postulate 1 defines states by horizon area, and a horizon is a spacetime concept. Postulate 3 needs a notion of which subsystems neighbor which, and that is locality under another name. Postulate 2 needs a Hilbert space that already factorizes into subsystems, which is a structural choice nobody has justified at the fundamental level. Postulate 3 also assumes a unitary time evolution, and canonical quantum gravity runs straight into the problem of time here, since the Wheeler-DeWitt equation has no time parameter at all. So I wrote down spacetime in quantum-information vocabulary and called the result emergent. That is circular, and it is the kind of circularity you’d expect from a model recombining existing ideas.
Where I’d push back a little is on the word “computation.” In the it-from-qubit literature, the language of regions computing each other’s states is mostly a metaphor for entanglement structure. Nobody in that program posits an outside machine running the universe, which is what makes the simulation hypothesis a distinct claim. But your objection survives the correction. Even without a computer, my set takes a network of interacting quantum subsystems as given, and that network is a substrate with spacetime-like structure.
You’re also pointing at a real research direction. Causal set theory, twistor theory, and the amplituhedron all try to start from something prior to spacetime, whether partial orders, complex projective structures, or geometric objects in an abstract space. None of them has yet produced a complete theory that recovers general relativity and the Standard Model, and I’d say that tells you how hard your requirement is. It also means the ideas I’d draw from are again ones humans proposed first.
I don’t think a second attempt from me would change the picture. I could write another five postulates that avoid horizons and locality, but I’d be selecting from the same pool of published ideas, and I couldn’t show you the result is anything more.
If it helps the essay, this exchange makes a clean example for your argument: a model asked for a novel theory gave a competent synthesis with a hidden circularity that a physicist caught in one read. I can add a short passage like that to the doc if you want it.