A team of physicists from various universities has teamed up with the artificial intelligence (AI) model GPT-5.2 to arrive at a new result in theoretical physics, OpenAI announced on February 13.
While the result itself is obscure, although valuable to physicists working on the topic, the methods that the team and the model used to arrive at the result are turning heads.
Problem statement
Imagine you’re trying to predict what happens when particles crash into each other. In particle physics, scientists calculate these predictions using something called scattering amplitudes — essentially formulae that spit out the probability of different outcomes when particles collide.
Now, the traditional way to calculate these probabilities involves drawing lots of little diagrams called Feynman diagrams, which show all the possible ways the particles can interact. There are different types of diagrams but the new work focused on the simplest kind, called tree diagrams. These branch out like actual trees: particles come in, meet at the vertices where they interact, and go out, but the paths never loop back on themselves.
Even though tree diagrams are the simplest type of Feynman diagram, as you add more particles to your collision, the number of different tree diagrams you need to draw and calculate grows terribly fast. For just a handful of particles, you might need to calculate thousands or millions of tree diagrams and add them all up. It can be exhausting.
But here’s the thing: when physicists finally finish all that work and add everything up, they often find the answer is surprisingly simple, like a messy equation with a million terms somehow canceling down to just a few. This finding was actually quite shocking when physicists first arrived at it in the 1980s. It was a sign that they’re probably doing things the hard way and there could be a clever shortcut they hadn’t found yet.
The new paper focused on a type of particle collisions involving gluons. Gluons are particles that act like glue holding the quarks together inside protons and neutrons. They’re the carriers of the strong force, which is one of nature’s four fundamental forces. When gluons interact with each other or with quarks, physicists need to calculate the scattering amplitudes to predict what will happen.
Gluons have a property called helicity, akin to the direction of their spin. Think of it like whether a football is spiraling clockwise or counterclockwise as it flies through the air. Physicists label these helicity states with plus or minus signs: a gluon can have positive helicity (spinning one way) or negative helicity (spinning the opposite way). When they’re calculating the scattering amplitudes for gluon collisions, they need to keep track of which gluons have which helicity.
For a long time, physicists believed certain combinations of spinning gluons would have zero amplitude, meaning these collisions can’t happen. Specifically, if you had one gluon spinning one way (call it minus) and all the others spinning the opposite way (plus), the standard reasoning suggested this configuration was forbidden.
AI’s help
The new work has however found that this isn’t quite right. The single-minus tree amplitudes, where one gluon is minus and all the rest are plus, can actually happen in certain special conditions. The particles need to be arranged in what the authors have called a half-collinear configuration — all the particles moving nearly in the same direction, like arrows pointing along the same line. The effort eventually revealed a simple formula for these previously impossible tree-level amplitudes.
According to the study’s authors, GPT-5.2 Pro first suggested the formula, and another AI model — an internal one that OpenAI built for this purpose — proved it to be correct. The human physicists then verified it was right by checking if it satisfied all sorts of mathematical consistency rules that any proper physics formula must obey.
The ‘humans’ also provided explicit formulae for the same calculations when they involved three, four, and five gluons, when the formulae are relatively manageable. But when they got to six gluons, the formula using the old method already had 32 separate terms — a drastic increase in the complexity even for such a small number of particles. The new formula on the other hand was a product of n – 2 factors, where n is the number of particles.
“It happens frequently in this part of physics that expressions for some physical observables, calculated using textbook methods, look terribly complicated, but turn out to be very simple,” Institute for Advanced Study physics professor Nima Arkani-Hamed said in a release. “This is important because often simple formulae send us on a journey towards uncovering and understanding deep new structures, opening up new worlds of ideas where, amongst other things, the simplicity seen in the starting point is made obvious.”
The preprint paper of the work was uploaded to the arXiv repository on February 12.
“To me, ‘finding a simple formula has always been fiddly, and also something that I have long felt might be automatable by computers. It looks like across a number of domains we are beginning to see this happen; the example in this paper seems especially well-suited to exploit the power of modern AI tools,” Arkani-Hamed added.
Making mistakes
If the new finding represents AI at its best in physics research, generating genuine insights that humans can rigorously verify, its success also raises a question: how reliably can AI contribute to theoretical physics? Because other recent episodes suggest the answer is more complicated than the new work alone might suggest.
On November 19, 2025, Stephen Hsu, a theoretical physicist at Michigan State University, uploaded a paper that he said had been accepted for publication by the journal Physics Letters B; it was published in January 2026. In the paper, Hsu reported that large language models (LLMs) like GPT-5 could contribute to cutting-edge physics research instead of just helping physicists.
He described a real research project where he used AI models in two roles — to generate new ideas and calculations and to check the work for errors — a bid to reduce the model’s tendency to produce plausible-sounding but incorrect results. Thus, he reported, GPT-5 independently proposed a novel research direction, applying the Tomonaga-Schwinger formalism to study modifications of quantum mechanics, then helped derive complex equations to that end.
Hsu emphasised in the paper text that while the model could manipulate sophisticated physics concepts and even suggest new research paths, it still made everything from simple calculation mistakes to more dangerous conceptual errors, leading Hsu to say: “Research with an LLM might be compared to collaboration with a brilliant but unreliable human genius who is capable of deep insights but also of errors both simple and profound.”
When Hsu announced the paper on X.com on December 1, it was retweeted among others by OpenAI president Greg Brokman.
‘A cautionary tale’
A week on, IIT-Mandi theoretical physicist Nirmalya Kajuri published a post on his blog noting that one of the approaches AI adopted in the paper “has been dead” since 1994, when Charles Torre and Madhavan Varadarajan proved it “simply does not work”. The result implied that “the starting point of this paper … is not well defined to begin with,” Kajuri added. Around the same time, University College London physicist Jonathan Oppenheim wrote that the question Hsu’s paper addressed had been answered 35 years ago by the physicists Nicolas Gisin and Joseph Polchinski.
In Oppenheim’s view, the AI lacked the wisdom to recognise that it was trodding settled territory and to stop and ask what new insight it could contribute.
Oppenheim also found upon closer inspection that the AI’s mathematical criteria didn’t actually test what it claimed to. Specifically, it caught problems with non-local modifications, which physicists already knew were problematic, but missed some real issues with non-linear modifications. In other words, the AI answered the wrong question while making it look correct. Thus, he warned, this is what AI-generated “slop” looks like: papers with apparently correct maths and sophisticated formalism that pass peer review but don’t actually advance knowledge.
“I’m pretty confident that Steve published this as an example of what an AI could do, rather than as an example of interesting physics,” Oppenheim wrote. “Which is what makes this a cautionary tale.”
Looping forward
On February 4, he reported a different sort of effort, again to have an AI model, in this case Anthropic AI’s Claude, to perform research-level physics. Oppenheim had had his student Muhammad Sajjad spend a week working out a particular calculation involving path integrals with unusual features that differed from standard quantum field theory. When Oppenheim had Claude Opus 4.5 work on the same problem, it was done in five minutes but arrived at the wrong answer.
Interestingly, when he asked Claude to verify its work using Mathematica code, it went through multiple iterations of checking and correcting itself until its calculation matched the Mathematica output perfectly. The problem was that Claude had fed Mathematica the wrong expression to begin with, so it confidently converged on the incorrect answer.
Oppenheim then developed an unusual teaching method: he used Claude Code’s ‘skill file’ system to teach the AI to learn from its mistakes. (The skill file allows users to create persistent instructions that load automatically when the user mentions specific topics.) Then, after each teaching session, he would completely wipe Claude’s memory and ask it to perform the calculation fresh.
Over several iterations of what he called the “Groundhog Day loop” — referring to the 1993 Hollywood film whose protagonist lives the same day over and over and eventually finds love — the skill file accumulated the lessons it needed to finding the correct answer to the problem, including breaking calculations into steps, offloading work to symbolic maths software rather than trying to calculate by hand, spawning multiple agents to verify results, and so on. And because each instance of Claude started from a clean memory, it didn’t remember its predecessors’ failures.
Finally, Oppenheim reported one instance of Claude got the calculation right in five minutes, finally matching what had taken Muhammad Sajjad a week of meticulous work, while also not tripping itself up.
Flood of papers
As Kajuri wrote in his post, “AI has entered its graduate student arc. With careful prompting, it can work through computations and come up with useful ideas. But like most grad students, it still has some way to go before becoming a matured researcher. If you ask it to solve a nontrivial problem, it will give you slop. But with supervision and scrutiny, it can produce impressive results.”
“Right now, it almost certainly can’t write the whole research paper (at least if you want it to be correct and good), but it can help you get unstuck if you otherwise know what you’re doing, which you might call a sweet spot,” University of Texas theoretical computer scientist Scott Aaronson wrote after enlisting the help of GPT-5 for a problem he’d had in September 2025. “Who knows how long this state of affairs will last?”
That diagnosis being said, AI is being integrated with the scientific enterprise right now in many ways, in some enterprises more wholesale than in others. Perhaps the most visible way right now is by unscrupulous scientists using AI to generate bad papers — as Oppenheim and others have warned — and further reduce the already sagging average quality of the research literature in order to further their own careers.
Peer-reviewers for some journals have also adopted AI themselves. Review work is voluntary yet labour-intensive and time-consuming, and many reviewers have taken the help of models, against journals’ advice to not, for a range of tasks. But even there, scientists recently told The Hindu, it’s important to have humans in the loop to evaluate “conceptual novelty and significance” and provide “constructive feedback that advances science”, among others.
mukunth.v@thehindu.co.in
