Quite profoundly—but in a way that is easy to misunderstand. AI has not yet given cosmology a new Einsteinian theory. Its present impact is more methodological: it is changing what cosmologists can extract from observations, what they can simulate, and increasingly what counts as “understanding.”
I would divide the impact into five levels.
1. AI lets cosmology use information that classical statistics throws away. Traditional cosmology compresses enormous datasets into a few statistics—power spectra, correlation functions, BAO scales, etc. Machine-learning methods can instead learn from the full non-linear morphology of the cosmic web: filaments, voids, halos, galaxy environments, weak-lensing maps. This is especially important because forthcoming surveys contain far more information than can comfortably be handled by hand-designed summary statistics. Simulation-based inference is becoming a major route from these complicated observations directly to posterior distributions for parameters such as \Omega_m, \sigma_8, neutrino mass and dark-energy parameters. Recent work already shows increasingly sophisticated ML emulators being inserted into conventional Bayesian/MCMC inference pipelines.
2. AI makes it possible to manufacture millions of plausible universes cheaply. This may ultimately be its largest practical effect. A high-resolution N-body or hydrodynamical simulation is immensely expensive. Neural emulators and generative models learn the mapping
\{\Omega_m,\Omega_b,H_0,n_s,\sigma_8,w,M_\nu,\ldots\}
\longrightarrow
\text{observable Universe}
and thereafter generate approximate predictions in fractions of the original computation time. Current emulators already predict nonlinear matter power spectra over multi-parameter cosmologies, and techniques such as meta-learning are making these models adaptable rather than tied to a single observing configuration.
This changes the epistemology of cosmology slightly. Instead of:
theory → equations → solve equations → compare with Universe,
we increasingly get:
theory → enormous ensemble of simulated universes → AI learns the manifold → compare our Universe with the manifold.
The AI is therefore becoming something like an interpolating layer between physical law and observation.
3. AI is particularly powerful where cosmology becomes an inverse problem. Cosmologists observe photons, redshifts, shear, gravitational waves and galaxy distributions, but they want to infer invisible causes: initial conditions, dark matter distribution, cosmological parameters, galaxy-formation histories. AI is exceptionally well suited to this many-to-one inversion.
Gravitational-wave astronomy gives a particularly clear example. ML methods can rapidly infer source parameters and sky positions, allowing conventional telescopes to turn toward an event quickly enough to catch its electromagnetic counterpart. Such multimessenger observations can in turn constrain cosmology—for instance through “standard sirens” as independent measurements of cosmic distance and H_0.
And this is coming at exactly the moment cosmology needs better inference. ΛCDM is extraordinarily successful, but current datasets continue to raise questions involving H_0, S_8, dark energy and structure formation. The 2025–26 cosmology literature still treats these tensions as unresolved rather than as evidence for one agreed replacement theory.
4. The really interesting change may be AI-assisted discovery rather than AI-assisted measurement. We are only beginning this stage. One can imagine an AI being asked not merely:
“What value of H_0 best fits these data?”
but:
“What compact mathematical law generates all these observations with the fewest assumptions?”
Symbolic regression, differentiable physics, neural operators and automated hypothesis generation could eventually search spaces of physical theories that no human group could inspect manually.
That would be a qualitative transition:
\text{AI as statistician}
\rightarrow
\text{AI as simulator}
\rightarrow
\text{AI as theorist}.
We are decisively in the first two stages. The third remains embryonic.
And here there is an important negative result: today’s LLMs are not autonomous cosmologists. A 2026 controlled study comparing AI and expert literature searches in physics, astrophysics and cosmology found surprisingly little overlap between the literature selected by humans and by mid-2025 AI systems, with substantial bibliographic errors in older models. So we should distinguish AI scientific capability from the fluent appearance of scientific understanding.
5. And this produces, to me, the deepest problem: cosmology may acquire prediction without comprehension. Nature Astronomy has begun discussing precisely this issue—what “understanding” means when important astronomical results emerge from highly capable but partly opaque machine-learning systems.
Suppose an AI discovers a mapping that predicts every galaxy survey to extraordinary precision but no cosmologist can translate the mapping into an intelligible physical theory.
Have we understood the Universe?
Or merely built an extraordinarily accurate oracle?
This is not a minor philosophical problem. Astronomy is already debating whether AI changes the meaning of scientific expertise itself—because coding, literature synthesis, modelling and eventually hypothesis generation can increasingly be delegated.
There is therefore a rather beautiful possibility. Cosmology may become the first science in which humanity possesses a map of reality substantially more detailed than its own conceptual understanding of that map.
And there is another implication that I think connects particularly strongly to our recent discussion of one → two → relation → no-relation. Classical cosmology begins from remarkably few human-made variables: H_0,\Omega_m,\Omega_\Lambda,\sigma_8,\ldots. We force an unimaginably complicated Universe into a small conceptual coordinate system.
AI may reverse that direction.
Instead of asking the Universe to become intelligible in our variables, we allow a machine to encounter an astronomical field consisting of billions of galaxies and discover its own high-dimensional latent coordinates.
Thus:
\boxed{\text{human cosmology: Universe }\rightarrow\text{ a few concepts}}
whereas increasingly,
\boxed{\text{AI cosmology: Universe }\rightarrow\text{ relations too numerous to name}.}
That may eventually prove more consequential than faster simulations. AI could push cosmology from a science primarily of human-comprehensible objects and parameters toward a science of relational structure.
And then the extraordinarily interesting question becomes not simply “Can AI discover the laws of the Universe?” but “Must a law of the Universe ultimately be intelligible to Homo sapiens in order to count as a law?”
That, I suspect, is where AI’s impact on cosmology becomes genuinely philosophical rather than computational.
