OpenAI and researchers at Princeton's Institute for Advanced Study announced they had used artificial intelligence to discover a new mathematical result in quantum gravity theory. The findings were released on the open-access scientific preprint repository arXiv.

The online outlet Gigazine reported on March 5 that the study used OpenAI's latest model, GPT-5.2 Pro. The researchers explained that the AI-proposed approach played an important role in finding a new mathematical structure.

The core of the study is analysing the "scattering amplitude" of the graviton, a hypothetical particle that mediates gravitational interactions. The team focused on the so-called "single-minus" amplitude structure, in which one particle has negative helicity and the remaining particles have positive helicity. Helicity is a physical quantity that indicates how a particle's spin direction aligns with its direction of motion and is an important factor that determines the characteristics of particle interactions.

Existing theory held that the single-minus amplitude should be zero at the "tree level", the most basic interaction stage. The researchers, however, proved that if particle momenta satisfy a specific arrangement condition called the "half-collinear regime", the amplitude need not be zero and can exist as a clear mathematical distribution. The result suggests a significant possible revision to the existing assumption.

The AI's role in the research process also drew attention. The researchers provided GPT-5.2 Pro with a previously published paper related to gluons as reference material and asked it to apply similar mathematical structures to quantum gravity theory. As a result, the AI not only quickly organised calculations that a human researcher would have to carry out over a long period, but also presented a new type of solution, the report said.

The team described the approach proposed by GPT-5.2 Pro as an "astonishingly elegant method". They also said the AI helped not only with calculations but also with drafting the paper.

OpenAI stressed that the study is an example showing AI can make a practical contribution to the process of scientific discovery. The researchers said the most time-consuming stage in the actual research process was not generating hypotheses but verifying the results the AI derived, checking consistency and organising them into a paper.

This is assessed as a change showing that while AI greatly increases speed in the early discovery stage of research, the role of human researchers is shifting to verification and interpretation of results.