©Anton Maksimov 5642.su via Insplash
There are statements that end up in textbooks and stay there seemingly forever. One of them essentially stated that a certain kind of interaction between gluons - the particles that hold the atomic nucleus together - simply cannot happen. But that turns out not to be true.
A new study, to which GPT-5.2 contributed, shows that that interaction is not impossible at all. It occurs under very specific conditions, but it exists. And that's enough to reopen an issue that had been considered dismissed.
The research, titled 'Single-minus gluon tree amplitudes are nonzero', has been published on arXiv and is currently being peer-reviewed. The authors include physicists from institutions such as the Institute for Advanced Study, Vanderbilt University, the University of Cambridge, Harvard and OpenAI.
What does this all mean?
When two particles collide, physicists don't just watch what happens: they calculate the probability that a certain type of interaction will occur. That calculation is based on a number called the scattering amplitude. Without these amplitudes, we wouldn't be able to predict what happens in particle accelerators or in high-energy phenomena in the early universe.
In the case of gluons, the carriers of the strong nuclear force that holds protons and neutrons together, many of these amplitudes are surprisingly simple if you consider only the most 'direct' interactions, with no additional quantum complications. This is what physicists call the 'tree' level. But there was one exception that stood as definitive. If one gluon has a certain spin configuration (called negative helicity) and all the others have the opposite (positive helicity), then the amplitude was set to zero. Freely translated: that interaction does not take place.
The study shows that this conclusion is true only if the particles move in a 'generic' way. However, there exists a special configuration, called the semi-collinear regime, in which the particles are aligned in a special way. In that case, the interaction is non-zero. This is not a trivial detail. It's like discovering that a door you thought was bricked shut opens anyway ... provided you approach it from exactly the right angle.
How GPT-5.2 helped find a formula that physicists had trouble discerning
This is where artificial intelligence comes in. The researchers had already done the calculations by hand for some specific cases. The problem? The formulas quickly became extremely long and almost unmanageable. The more particles were added, the more explosive the complexity increased.
GPT-5.2 Pro took those complicated expressions and simplified them. But above all, it discovered a recurring pattern and managed to propose a general formula valid for any number of particles. It was not a 'lucky hunch'. An internal version of the model worked for about 12 hours, reconstructing step by step the mathematical reasoning until it arrived at the same formula and gave a formal proof for it.
The validity of the result was then tested using the standard methods of theoretical physics, including the Berends-Giele recursion relation and the so-called soft theorem, which imposes very precise rules on the behavior of interactions at low energy. In other words, this is not just an AI's hunch. It is computed and probed mathematics.
AI and science: collaboration, not replacement
Theoretical physicist Nima Arkani-Hamed reacted enthusiastically to the appearance of such simple expressions in a field known for its complexity. Often, he recalled, formulas that seem unmanageable by traditional methods turn out to be extraordinarily elegant once you find the right key. And precisely the search for these kinds of simple structures could be one of the most promising areas for intelligent automation.
Nathaniel Craig also emphasized that this work represents research at a high academic level and provides a concrete model for collaboration between physicists and large language models. The question at this point is no longer whether artificial intelligence will enter the theoretical laboratories, but how the way of doing science will change once the dialogue between humans and AI becomes structural.
At a time when technological innovation is often described almost exclusively in economic or commercial terms, this discovery reminds us of something more fundamental: AI can also help us better understand the universe. And perhaps, occasionally, to question what we thought we knew for sure.
(©OpenAI via GreenMe.it 2026/Managing Editor: Selma Keshkire - The Press Junction/Picture: ©picture alliance/dpa/dpa-POOL | Peter Kneffel)