OpenAI Preprint Shows Graviton Amplitudes Once Assumed Zero

A new preprint from researchers working on behalf of OpenAI shows that a class of graviton interactions long assumed to vanish can in fact arise under well-defined kinematic conditions. The paper, titled "Single-minus graviton tree amplitudes are nonzero," is authored by Alfredo Guevara (Institute for Advanced Study), Alexandru Lupsasca (Vanderbilt University and OpenAI), David Skinner (University of Cambridge), Andrew Strominger (Harvard University), and Kevin Weil (OpenAI). It extends recent results the group obtained for gluons to the gravitational setting, and OpenAI has published the preprint as a PDF while inviting feedback from the community.

The stakes are high in two directions at once. The result touches the central open problem of reconciling quantum mechanics with Einstein's general relativity, and it arrives through an AI-assisted workflow in which GPT-5.2 Pro both solved the core problem and drafted the paper.

What the paper actually shows

Scattering amplitudes are the mathematical quantities physicists use to calculate the probability that particles interact in particular ways. Rather than tracking every intermediate step of a collision through many diagrams, amplitudes encode final observable outcomes in compact form. Over several decades, researchers have found that amplitudes often display unexpected simplicity, revealing hidden mathematical structure not obvious from traditional calculations.

The preprint studies gravitons, the quantum particles associated with gravity in quantum field theory. The authors analyze a configuration known as a single-minus amplitude: one particle carries negative helicity while all remaining particles carry positive helicity. Helicity describes the orientation of a particle's spin relative to its direction of motion, and it plays an important role in determining how interactions occur.

Standard textbook arguments suggest these amplitudes should vanish at tree level, the simplest level of approximation, where only the most direct interaction diagrams are considered and quantum loop effects are ignored.

The preprint breaks that assumption in a precise way. The vanishing conclusion depends on assuming generic particle motion. When particle momenta satisfy a special alignment known as the half-collinear regime, the usual argument no longer applies. In this regime, the amplitudes do not vanish. Instead, they exist as well-defined mathematical distributions supported on a restricted region of momentum space. The authors derive explicit formulas describing these interactions and show the formulas follow from symmetry principles and from recursion relations that build complex interactions from simpler ones.

A fifty-year-old symmetry, realized

The result is a small step toward reconciling quantum mechanics with general relativity, according to the authors. The single-minus amplitudes realize an infinite-dimensional "w-(1+∞)" symmetry. Penrose discovered this symmetry a half century ago in the context of classical gravity, and many researchers expect it to play a central role in quantizing the gravitational field. The preprint shows how, in the simplest possible context, this symmetry acts on gravitons, the elementary quantum bits of the gravitational field.

The final formulas were verified analytically and checked for consistency with known physical limits. After further interaction with GPT-5.2 Pro, the amplitudes were also found to be consistent with this infinite-dimensional symmetry first studied in connection with gravity by Roger Penrose.

How GPT-5.2 Pro produced the construction

The methodology section is the part of the preprint most likely to draw attention from anyone tracking AI in research.

Gravity and gauge theory share deep conceptual relationships, but their calculations differ substantially in practice. The earlier gluon result demonstrated that a previously neglected helicity configuration could produce nonzero amplitudes under special conditions. After that work was completed, the gluon paper was provided to GPT-5.2 Pro as context. Using it as a reference point, the model was asked to construct the corresponding amplitudes in quantum gravity, an extension the authors say would have taken human authors considerable time to derive.

GPT-5.2 Pro solved the problem using a technique the authors describe as beautiful and surprising: the directed matrix-tree theorem. It also produced what the authors call an excellent preliminary draft of the paper. OpenAI has published a transcript of the initial exchange between the researchers and the model.

The derivation combines several established tools in amplitude theory. Recursion relations iteratively construct many-particle interactions from smaller building blocks. Symmetry constraints restrict the allowed form of the result. The gravitational construction produced through this process was subsequently proven using standard analytic methods.

The transition from gluons to gravitons illustrates how mathematical insight can transfer across neighboring areas of theoretical physics. The two theories describe different fundamental forces, but they share structural features that allow ideas developed in one setting to inform the other. Providing the gluon result as an anchor enabled exploration of this connection.

The bottleneck is shifting to verification

An observation the authors draw from this and related projects concerns the pace of discovery, and it may be the most consequential claim in the announcement.

For this project, most of the time elapsed since the previous gluon result went to confirming derivations, checking consistency, and preparing formal write-ups rather than generating initial conjectures. The authors describe this sequence of results as a significant shift, with verification and exposition representing the dominant share of effort.

That framing matters for how the field reads the result. If AI systems increasingly handle conjecture and first-pass derivation, the human workload concentrates on checking and exposition, the parts of research where conventional standards of mathematical verification and scientific rigor live. The authors themselves frame the work this way: the preprint contributes to an ongoing effort to understand how AI-assisted reasoning can participate in theoretical research while maintaining those standards.

The names on the paper give the claim weight. Andrew Strominger of Harvard is a leading figure in gravitational scattering and asymptotic symmetries. David Skinner of Cambridge and Alfredo Guevara of the Institute for Advanced Study work directly in amplitude theory. Lupsasca holds a joint position at Vanderbilt and OpenAI, and Weil is at OpenAI.

Further extensions of these results are currently under investigation. The immediate question for the amplitudes community is whether the half-collinear graviton formulas hold up under independent scrutiny, and whether the w-(1+∞) connection opens a tractable route to quantizing gravity's simplest observables.