GPT-5.2 AI proposes a result in physics

GPT-5.2 Pro has proposed a formula for a type of interaction between particles that theoretical physics once considered impossible. Researchers from OpenAI and several universities have demonstrated and verified it in a new preprint, although the work is still awaiting peer review.

The study focuses on gluons, the particles that transmit the strong nuclear force, which holds together the components of atomic nuclei. The result does not claim that this interaction occurs in every situation: it appears under a very specific configuration of the particles.

An interaction that was not exactly impossible

To calculate what can happen when particles collide, physicists use a tool called a scattering amplitude. It is not the final probability, but the mathematical quantity used to calculate that probability.

In the case analyzed, one gluon has a spin orientation called negative helicity, while all the others have positive helicity. Standard arguments indicated that the corresponding amplitude should be zero, meaning that the interaction could not occur at the simplest level of the calculation.

The new preprint shows that this conclusion was too broad. The earlier argument works when the particles have generic momenta, but it no longer applies when their momenta, a combination of direction and energy, satisfy a special alignment known as the semicollinear regime.

In that specific region, the amplitude is not zero. The team calculated its value and found an expression that was much simpler than traditional calculations suggested.

What GPT-5.2 actually did

The authors first calculated several small cases by hand using Feynman diagrams, the standard representations of possible interactions between particles. The number of terms grew very quickly as the number of gluons increased.

GPT-5.2 Pro simplified those expressions, detected a pattern and proposed a general formula. An internal version of the model then worked for 12 hours to arrive at the same formula again and construct a formal proof.

The validation was not left to the model. The researchers analytically checked that the formula satisfies the Berends-Giele relation, a standard procedure for assembling complex interactions from smaller ones. They also compared it with a soft theorem, which indicates how the result should behave when one of the particles carries energy close to zero.

That order matters: the AI suggested the structure, but expert physicists checked that the proposal fit the known rules of quantum field theory.

Why this advance matters to you

It does not mean that AI has replaced a laboratory or discovered a new particle. The result is mathematical and concerns a specialized problem in theoretical physics. Its value lies in showing a different form of collaboration: finding patterns that are difficult to see in calculations that grow disproportionately, then subjecting them to independent tests.

The work also opens new lines of research. One of them is calculating equivalent interactions for gravitons, hypothetical particles associated with gravity. OpenAI says those extensions have already been carried out with help from GPT-5.2, although their details will be published separately.

The document is available as a preprint on arXiv and has been submitted for publication. Until other researchers review it, it should be treated as a promising result, not a definitive conclusion.

What matters is not only that a model found a formula. It is that AI is beginning to prove useful for a specific, verifiable scientific task: reducing unmanageable calculations, detecting regularities and proposing conjectures that experts can prove or reject. The next step will be to determine whether this method works reliably on other problems, beyond this particularly favorable case.