On 13 February 2026, OpenAI published a post titled “GPT-5.2 derives a new result in theoretical physics”, alongside a preprint on arXiv. Headlines quickly turned it into “AI solves decades-old physics puzzle”. The real story is more specific and, in some ways, more interesting: a group of well-known physicists used GPT-5.2 Pro to spot a pattern they hadn’t found by hand, and then used a version of the model to help prove it.
Here’s what the paper claims, what the AI actually contributed, and why the caveats matter as much as the result.
The result in plain terms
The preprint, “Single-minus gluon tree amplitudes are nonzero” (arXiv:2602.12176), is about scattering amplitudes: the mathematical objects physicists use to calculate how particles interact. Gluons are the particles that carry the strong nuclear force, and each gluon in a scattering process has a “helicity”, which can be thought of as plus or minus.
A standard result taught in quantum field theory is that tree-level amplitudes where only one gluon has negative helicity (the “single-minus” case) vanish. The paper shows that this isn’t always true. In a specific, carefully defined kinematic setup, called the half-collinear regime, these amplitudes can be nonzero, and the authors give a closed-form formula for them that works for any number of gluons.
Who did the work
According to OpenAI’s post, the human authors are Alfredo Guevara (Institute for Advanced Study), Alex Lupsasca (Vanderbilt University and OpenAI), David Skinner (University of Cambridge), Andrew Strominger (Harvard University) and Kevin Weil (OpenAI). These are serious researchers in the field, and the framing of the problem, the physical setup and the interpretation of the result came from them.
What the AI actually contributed
The division of labour, as OpenAI describes it, went roughly like this:
- Humans did the hard setup. The physicists computed the amplitudes by hand for up to six gluons. Those expressions grow very quickly in complexity as the number of particles increases.
- GPT-5.2 Pro simplified and generalised. Given those results, the model simplified the expressions and proposed a single formula that it conjectured would hold for any number of gluons.
- A scaffolded version worked on the proof. An internal, scaffolded version of GPT-5.2 then spent about 12 hours developing a formal proof of the formula.
- The result was checked independently. The authors verified the formula against known consistency conditions: the Berends-Giele recursion relation and constraints from soft theorems.
David Louapre, a physicist who wrote a careful analysis of the paper for the Hugging Face blog, singled out the conjecture step as the most interesting part: seeing a general closed-form pattern from a handful of special cases is the kind of insight that usually comes from an experienced researcher. He described it as a good use of AI that goes beyond brute-force symbolic manipulation.
The caveats
The result is real, but narrower than many headlines suggested. Louapre’s analysis and the paper itself point to several limits:
- It isn’t our everyday spacetime. The nonzero amplitudes appear in a setting with a different spacetime signature, often called (2,2) or “Klein” signature. Physicists use such settings as mathematical tools, but they aren’t the physical universe we measure directly.
- It’s a special slice of possibilities. The half-collinear regime is a precisely defined but non-generic set of particle momenta. The textbook statement still holds in the usual cases.
- The loophole wasn’t entirely unknown. Louapre notes that the possibility of nonzero single-minus amplitudes in this kind of setting had been hinted at in earlier theoretical work, but not worked out.
- The physical implications are unclear. It’s too early to say whether the result leads to new predictions or mainly deepens understanding of the mathematical structure of gauge theory.
- Details of the AI process are limited. OpenAI hasn’t published full details of the prompting and scaffolding, which makes it hard for others to judge exactly how much the model did on its own.
It’s also worth being precise about status: this is a preprint submitted for publication, not yet a peer-reviewed paper, at least at the time of the announcement.
Why it still matters
Even with those caveats, this is a meaningful example of how AI models can be useful in research. The model didn’t choose the problem, design the physical setup or decide what the result means. What it did was the step between examples and a general law: noticing the pattern, stating it compactly and helping prove it. That’s a large part of how theoretical progress often happens, and it’s traditionally the part that requires the most experience.
The most realistic reading is not “AI replaces physicists” but “a strong model can be a useful collaborator for experts”. The physicists here knew which calculations to do, could recognise a promising conjecture and could check the proof. Without that expertise, the model’s output would have been much harder to trust or use.
What developers and researchers can take from it
- AI is strongest with structured input. The model worked from carefully prepared examples. The same applies to code and data work: good inputs and clear examples produce much better results.
- Verification is non-negotiable. The authors checked the formula against independent constraints. Any AI-generated result, from a proof to a database migration, needs checks that don’t depend on the model.
- Be sceptical of headlines. “Solved a decades-old puzzle” and “found a nonzero amplitude in a special regime” are very different claims. Reading the primary source, here the OpenAI post and the arXiv paper, is the only reliable way to tell which one you’re looking at.
Summary
GPT-5.2 Pro helped a team of leading physicists find and prove a general formula showing that single-minus gluon tree amplitudes can be nonzero in a specific, non-physical kinematic regime. It’s a genuine research contribution and a good example of AI as a research collaborator, but it’s narrower than “AI solved physics”, and the human experts did the framing, checking and interpretation.
Sources: OpenAI: GPT-5.2 derives a new result in theoretical physics, arXiv:2602.12176, David Louapre, Hugging Face blog.
