OpenAI: GPT‑5.2 Derives And Proves New Formula In AI’s First Physics Breakthrough

In Brief

OpenAI’s latest preprint reports that GPT‑5.2 derived and proved a previously overlooked gluon amplitude, prompting renewed debate over whether advanced AI systems are now contributing original insights to theoretical physics.

OpenAI announced that it has released a new research preprint detailing work in which its GPT‑5.2 model independently identified a mathematical pattern and produced a formal proof, an outcome the organization describes as the first original theoretical physics contribution generated by one of its systems.

The study examines a long‑standing assumption in particle physics regarding gluon interactions, concluding that a class of scattering amplitudes previously believed to vanish can, under specific momentum conditions, in fact be nonzero.

The preprint, titled “Single-minus gluon tree amplitudes are nonzero,” is authored by researchers from the Institute for Advanced Study, Vanderbilt University, the University of Cambridge, Harvard University, and OpenAI. It focuses on scattering amplitudes, the quantities used to calculate the likelihood of particle interactions.

While many gluon amplitudes simplify at tree level, configurations involving one negative‑helicity gluon and multiple positive‑helicity gluons have traditionally been treated as yielding zero amplitude based on standard arguments.

The authors report that this conclusion does not hold in a precisely defined region of momentum space known as the half‑collinear regime, where particle momenta align in a special but mathematically consistent way. Within this regime, the amplitude does not vanish, and the team provides an explicit calculation. The finding opens new avenues for further work, including extensions to graviton amplitudes.

A notable aspect of the research concerns methodology. GPT‑5.2 Pro first proposed the general formula that appears as Eq. (39) in the preprint, after simplifying complex expressions derived manually for lower‑order cases. A scaffolded internal version of GPT‑5.2 then spent roughly 12 hours reasoning through the problem, independently arriving at the same formula and generating a formal proof. The result was subsequently verified using established techniques such as the Berends–Giele recursion relation and soft‑limit checks.

According to the authors, the approach has already been applied to extend the analysis from gluons to gravitons, with additional generalizations underway. OpenAI states that further AI‑assisted findings will be detailed in future publications.

Growing Evidence Of AI‑Led Discovery Fuels Debate Over Whether Machines Can Generate New Science

In the wake of OpenAI’s latest research milestone, debate is expected to continue over whether artificial intelligence can genuinely originate new scientific ideas. Skeptics are likely to question whether models are uncovering novel insights or simply recombining existing information in sophisticated ways.

However, the growing body of results emerging from advanced systems is making that distinction increasingly difficult to draw. As AI begins to probe and challenge assumptions that have shaped major scientific disciplines for decades, the notion of machine‑generated discovery is shifting from speculative fiction to a development that appears increasingly imminent.