An unlikely contributor has entered the world of high-energy theoretical physics: ChatGPT.
For decades, physicists believed that a particular interaction involving gluons – the massless particles that carry the strong nuclear force – could never occur. Now, researchers say OpenAI’s latest public model, ChatGPT-5.2 Pro, helped demonstrate that the process is in fact possible, deep within the complex internal structure of protons and neutrons.
The findings were presented last week at the annual meeting of the American Association for the Advancement of Science(AAAS), which publishes Science.
“The ideas are not revolutionary,” said Zvi Bern of the Mani L. Bhaumik Institute for Theoretical Physics at UCLA. “But what is revolutionary is that a machine can do this.”
A 40-Year Puzzle
Gluons bind quarks together to form protons and neutrons, and also bind protons and neutrons into atomic nuclei. The mathematics describing their interactions, known as scattering amplitudes, is notoriously complex.
In simple gluon collisions, physicists long believed that at least two particles had to possess negative helicity (a type of spin orientation). If only one gluon had negative helicity, the scattering amplitude was assumed to be zero , meaning the interaction could not happen.
About a year ago, three theorists identified a loophole: a single negative-helicity gluon might interact with positive-helicity gluons if all particles were traveling in roughly the same direction. Proving it, however, required navigating pages of unwieldy equations.
Andrew Strominger of Harvard University and his collaborators initially thought the calculation would take weeks. Instead, it stretched on for months. Alfredo Guevara of the Institute for Advanced Study eventually discovered a pattern in the equations, but generalizing the result for any number of gluons produced an expression dozens of terms long – too cumbersome to use.
The team suspected a clean, elegant formula was hidden inside the mess. They just couldn’t extract it.
Enter ChatGPT
At the same time, Alex Lupsasca of Vanderbilt University had joined OpenAI’s newly launched OpenAI for Science initiative. After reconnecting with Strominger, his former adviser, he suggested using ChatGPT as a test case.
The researchers fed the complex four-gluon expression into ChatGPT-5.2 Pro. Within about 20 minutes, the model simplified it. They repeated the process for five gluons, then six. In one case, the AI reduced a 32-term expression into a compact product spanning a single line.
Finally, they asked it to generalize the formula for any number of gluons. The system responded within minutes with what it described as an “obvious” generalized expression.
Concerned about possible hallucinations, the team rigorously checked the result. They found no errors.
“All of a sudden, I felt like my machine turned from a machine into a live being,” Strominger said.
To further validate the result, the team submitted the generalized formula to an internal OpenAI research model under development, nicknamed “SuperChat.” After roughly 12 hours of processing, the internal model produced a detailed proof that passed human scrutiny.
A Paradigm Shift?
The paper, posted to arXiv on 12 February, quickly gained attention online and sparked surprise at the AAAS meeting.
“What the OpenAI agent was able to do is impressive,” said Aida El-Khadra of the University of Illinois Urbana-Champaign.
The researchers believe the development could mark a turning point in how theoretical physics is conducted. Guevara suggested AI might soon become as integral to physics as it has to programming – handling routine derivations, checking errors, and accelerating research workflows.
The broader physics community has responded with cautious optimism. While many see AI as a powerful assistant for verification, drafting, and cross-disciplinary synthesis, concerns remain about transparency, training of graduate students, and overreliance on automated systems.
Still, few believe scientists are at risk of being replaced.
“None of this feels to me like scientists will be replaced,” El-Khadra said.
Lupsasca is already looking ahead. He hopes similar techniques could be applied to gravitons – hypothetical quantum particles that mediate gravity – and perhaps even help tackle one of physics’ greatest unsolved problems: reconciling quantum mechanics with gravity.
For now, the result stands as a striking milestone: after 40 years of near-intractable algebra, a large language model helped physicists uncover an elegant formula describing interactions among fundamental massless particles and proved that AI can meaningfully contribute to front-line theoretical research.
