AI as Co-Scientist:
From Black Hole Symmetry to
Particle Physics — The New Era
of Machine Insight

The LHC Rediscovers
Einstein's Symmetries

While the gluon amplitude result was the most dramatic single moment in AI-assisted physics in early 2026, a quieter but arguably more important line of work had been building for longer. At the University of California, San Diego, computer scientist Rose Yu and collaborators trained machine learning models to search for symmetries hidden inside raw experimental data from the Large Hadron Collider — with no prior knowledge of physics built into the training.

A symmetry in physics is a transformation that leaves the laws of nature unchanged. The most fundamental symmetries in the universe are Lorentz symmetries — the mathematical invariances that underpin Einstein's special theory of relativity. They encode the fact that the laws of physics are the same for all observers, regardless of their speed or direction. These symmetries are not just elegant; they are the architecture of reality as we understand it.

Yu's team applied their technique to LHC collision data and asked: can a model, without being told what to look for, discover Lorentz symmetry purely from data? The answer was yes. "We showed that, without knowing any physics, the model can discover the Lorentz symmetry purely from data," Yu said. The result is a proof of principle for something profound: AI can find the mathematical structures underlying physical law from raw observation, the same way a physicist would — but starting with no prior conceptual framework.

The implications extend well beyond confirming known symmetries. The same technique could, in principle, surface symmetries we do not yet know about — violations of Lorentz invariance, hidden conservation laws, patterns in high-energy collision data that point toward physics beyond the Standard Model. At the LHC, where tens of millions of collisions occur per second and only a tiny fraction of events are ever stored, AI is already making the first-pass decisions about what gets saved. The next step is asking it to find the patterns we did not think to search for.

AI-Newton:
Deriving Laws from Raw Data

The most conceptually ambitious frontier in AI-assisted physics is not amplitudes or symmetries — it is the question of whether a machine can, given only raw data, derive the laws of physics from scratch. In 2025, a Chinese research team published AI-Newton, a system that, after being fed experimental data, can autonomously discover key physics principles — including Newton's second law of motion — without any prior knowledge of mechanics being encoded in its training.

The system works by learning to break complex physical relationships into simpler sub-components, drawing on known, well-established equations as building blocks. Published in Nature, the work was immediately met with a nuanced response. Experts noted, correctly, that AI-Newton rediscovers known laws — it does not generate novel ones. Its reasoning is pattern extraction, not the kind of conceptual leap that produced general relativity or quantum mechanics. "Any AI, as they're designed today, is extremely unlikely to achieve the kind of breakthroughs required to come up with new physics," one Nobel Laureate physicist wrote, arguing that such discoveries require genuine creativity beyond optimisation.

But a complementary development — the PhyE2E framework, published in Nature Machine Intelligence in late 2025 — pushes the boundary further. PhyE2E derives compact, unit-consistent symbolic equations from raw space physics data using a transformer model and a divide-and-conquer approach. Crucially, it imposes physical plausibility constraints on its generated formulas, ensuring they respect dimensional consistency — something pure data-fitting approaches routinely violate.

Neither AI-Newton nor PhyE2E is yet generating physics that surprises physicists. What they represent is a critical capability being assembled: the ability to compress raw observational data into interpretable symbolic laws. As telescopes like Vera Rubin Observatory produce petabytes of sky survey data nightly, and particle detectors record millions of collisions per second, the physics of the universe increasingly lives in datasets too large for human intuition to navigate unaided.

What AI Can —
And Cannot — Do in Physics

The gluon result has generated genuine excitement and an equal measure of careful qualification. The physicists themselves are precise about what the AI did: it identified a pattern, simplified a formula, and verified a proof. It did not propose the underlying physical hypothesis — that was Guevara, Skinner, and Strominger. It did not design the experiment. It did not know what a gluon was until told. "The ideas are not revolutionary," said Zvi Bern, a particle theorist at UCLA. "But what is revolutionary is that a machine can do this."

The distinction matters. What AI has demonstrated, in multiple physics contexts now, is an extraordinary ability to perform the most technically demanding algebraic and pattern-recognition steps in a physicist's workflow — the steps that currently consume years of postdoctoral effort and occasionally produce careers-worth of dead ends. What it has not yet demonstrated is the ability to ask the right question in the first place.

Physics has always distinguished between two kinds of intellectual work: calculation and understanding. Feynman's famous dictum — that he could not truly understand something until he had built a simpler model of it — points at the gap. AI is now extraordinarily capable at the calculation end. The question of whether it can genuinely understand — whether it can look at a formula it produced and say why nature arranged itself this way — remains open, contested, and deeply philosophically interesting.

The Road Ahead:
Five Frontiers That Matter

• 01

  Beyond the Standard Model — AI anomaly hunting at the LHC

  The Large Hadron Collider produces data at a scale no human team can fully analyse. The next phase of unsupervised AI deployment at CERN is aimed not at confirming known physics, but at flagging deviations from it — events that do not fit the Standard Model's predictions. The discovery of new physics, when it comes, may arrive as an anomaly that an AI flagged and a human chose to follow up. The question is not whether the AI will find something unusual. It is whether physicists will know how to ask the right follow-up question.

• 02

  Gravitational wave astronomy — the next generation

  LIGO and its successor detectors are already deploying AI to process gravitational wave signals in real time. The AI-designed 3km resonator loop — unintuitive to every human physicist who saw it — demonstrated that there is unexplored design space in the instruments themselves. Next-generation detectors will almost certainly be designed in collaboration with AI optimisation tools, and their analysis pipelines may increasingly be driven by models that flag events no human specification anticipated.

• 03

  From proof verification to proof generation — the mathematical frontier

  Gemini Deep Think's performance at International Mathematics Olympiad gold-medal standard, and its subsequent progress toward PhD-level exercises, is closing the gap between AI as proof-checker and AI as proof-generator. Theoretical physics and mathematics share a deep boundary — the gluon amplitude result is as much a mathematical result as a physical one. As models like Aletheia push further into research-level mathematics, the frontier between "AI checks the calculation" and "AI does the calculation" will dissolve.

• 04

  The interpretability problem — building the physics story

  The most important unresolved challenge in AI-assisted physics is not capability — it is understanding. AI's dark matter density formula "describes the data very well but is lacking the story about how you get there." LIGO's AI-designed resonator loop required months of analysis to explain. Physics does not simply want correct answers; it wants explanations that can be built upon, taught, and extended by human intuition. The next generation of physics AI will need interpretability as a design principle, not an afterthought.

• 05

  Retooling, not retiring — what physics education must become

  Strominger was asked whether AI would make him obsolete. "Call it vanity. I think I'm irreplaceable," he said. "I think it will empower us to do more, but we have to retool. Good scientists have to retool all the time." The generation of physicists currently in graduate school will work alongside AI collaborators for their entire careers. Physics education has not yet fully reckoned with what that means: which skills to deepen, which calculations to automate, and how to train human intuition to ask the questions that machines cannot yet form on their own.

A Timeline of
Machine Insight in Physics