An unexpected breakthrough in particle physics has emerged from an unlikely source: ChatGPT. For years, physicists believed a certain interaction involving gluons—the particles that mediate the strong nuclear force—was impossible. But recent work led by researchers including Andrew Strominger of Harvard University and Alfredo Guevara of the Institute for Advanced Study has shown that such a process can occur, albeit in highly complex conditions deep within protons and neutrons. The findings were presented at the annual meeting of the American Association for the Advancement of Science and have since gained widespread attention. Gluons, which bind quarks into protons and neutrons, carry a property called helicity—essentially, the direction of their spin relative to their motion. For decades, it was thought that in any gluon collision, at least two gluons must have negative helicity for the interaction to be possible. If only one had negative helicity, the mathematical probability—known as a scattering amplitude—was believed to be zero. However, a year ago, a small group of theorists identified a potential exception: a scenario where a single negative-helicity gluon interacts with others of positive helicity, provided all particles move in nearly the same direction. Despite months of manual calculations, the team struggled to prove the idea. The mathematical expressions were overly complex, with dozens of terms that defied simplification. They suspected a more elegant formula existed—similar to one discovered in the 1980s for a related interaction—but couldn’t uncover it. Enter Alex Lupsasca, a theoretical physicist who joined OpenAI for Science and was tasked with enhancing ChatGPT’s scientific capabilities. He connected with Strominger and proposed testing the model on this very problem. Initially skeptical, Lupsasca asked the latest public version of the model, ChatGPT-5.2 Pro, to simplify the scattering amplitude for four gluons. Within minutes, it reduced a complex sum of 32 terms into a single, clean line. The model then tackled five and six gluons, again delivering simplified results. When asked to generalize the formula for any number of gluons, it responded in under two minutes with a concise, elegant expression it labeled “obvious.” The researchers, fearing a hallucination, rigorously checked the result—and found it correct. To verify the solution further, they fed the formula into a more advanced, internal OpenAI model known only as “SuperChat.” After 12 hours of processing, it generated a full mathematical proof that passed human scrutiny. The team published the results on arXiv on February 12, and the paper quickly went viral. At the AAAS meeting, physicists were stunned. “What the OpenAI agent was able to do is impressive,” said Aida El-Khadra of the University of Illinois Urbana-Champaign. While concerns remain about AI’s role in academia—such as transparency and the potential impact on graduate training—experts agree that AI is not replacing scientists. Instead, it’s becoming a powerful tool for accelerating discovery. Zvi Bern of UCLA, while cautious, called the result “revolutionary” in terms of what machines can now contribute. The team is now exploring whether similar AI-driven insights can help solve one of physics’ greatest challenges: unifying quantum mechanics and gravity. Lupsasca plans to apply the same method to gravitons, the hypothetical particles of gravity, aiming to develop a well-behaved quantum theory of gravity—possibly by the end of the year.