AI helps compute graviton amplitudes in quantum gravity

Theoretical physics has long been regarded as a preserve of the human mind, a field where intuition and years of specialized training are irreplaceable. A new preprint from a group of researchers challenges that axiom: the large language model GPT-5.2 Pro helped, for the first time, to derive and verify nonzero tree-level graviton amplitudes in quantum gravity. This is not just a technical curiosity. It signals that artificial intelligence is beginning to do real work in a domain where, until now, only a narrow circle of specialists could operate.

To grasp why the result matters, some context helps. Scattering amplitudes are mathematical objects that describe the probabilities of elementary particle interactions. In quantum field theory, computing them with classical Feynman diagrams is notoriously laborious: the number of terms grows factorially with the number of particles involved. The so-called single-minus amplitude method, which covers a special class of helicity configurations in which exactly one of the external momenta has negative helicity, significantly cuts down the analytical work. Until recently, however, the method was applied mainly to gluons in gauge theories such as quantum chromodynamics. Extending the approach to gravitons, the particles that carry the gravitational interaction in quantum gravity, was a separate and nontrivial problem.

Gravitons are, in theoretical terms, considerably more complex than gluons. Their scattering amplitudes, even at the level of tree diagrams with no quantum corrections, produce expressions of truly daunting bulk. This is where GPT-5.

2 Pro came in. The researchers used the model not as a replacement for a human physicist but as an analytical assistant able to handle symbolic computation, follow long algebraic chains and check intermediate results for correctness. The model helped derive specific nonzero amplitudes and verified them by comparing them against known relations, in particular the Kawabata-Luforo-Stevens relations, which link gravitational and gauge amplitudes.

It is important to stress that this was not numerical simulation or parameter fitting: GPT-5.2 Pro worked with symbolic algebra, the kind that traditionally requires a professionally qualified human.

This distinction is fundamental. In recent years, neural networks have been widely applied in physics to solve differential equations, optimize molecular structures or speed up simulations. But those are all tasks where the model operates on numbers or approximates functions. Analytical calculation in theoretical physics is a fundamentally different matter: it requires manipulating symbolic expressions and understanding the structure of Lie algebras, the symmetry group of the problem and the physical constraints imposed on the result. That a language model can handle such work at a level sufficient for inclusion in a scientific preprint marks a qualitative shift in what the technology can do.

For industry and the academic community, the result opens several important prospects. First, it legitimizes the use of LLMs in fundamental theoretical science, a field where such tools were previously viewed with skepticism. Second, it points to a concrete niche where language models deliver real gains: not replacing the scientist, but speeding up the most routine, if technically demanding, stages of the work. Cumbersome algebraic manipulations that can take a theoretical physicist weeks could potentially shrink to hours. Third, the publication stimulates a debate about verification: how should the community treat results obtained partly with the help of AI? The authors, by all appearances, bet on transparency by explicitly stating the role of GPT-5.2 Pro in the work.

Quantum gravity remains one of the main unsolved problems of fundamental physics. The theory still has no experimental confirmation, and the analytical apparatus needed to develop it is extremely complex. If language models can take on part of that analytical load, the pace of theoretical research may increase noticeably. Not because AI understands physics in the human sense, but because it can operate with great precision and speed on the formal structures through which physics expresses itself. This is a preprint worth following: if the results pass peer review, we may be witnessing the start of a new working practice in theoretical physics.