The AI simplified the cumbersome expressions and suggested a general formula, which was then verified manually.
It would seem that a chatbot is a far cry from theoretical physics. But in February 2026, physicists were seriously discussing a result ChatGPT had led them to at an AAAS meeting. It concerned an "impossible" interaction between gluons, the particles that bind quarks into protons and neutrons. For decades, it was believed that such a process simply didn't occur at the most basic level. Now it turns out that it's possible, but it's hidden under very specific conditions, similar to what occurs in the dense and chaotic interior of nucleons.
Gluons mediate the strong force. It's so powerful and nonlinear that even "simple" gluon collisions are described by formulas that quickly give you a headache. Physicists reduce each such collision to a scattering amplitude, a mathematical expression that allows you to calculate the probability of the process. The problem is that these amplitudes grow rapidly, and for just a few particles, it's difficult to simplify them, let alone even accurately finalize them.
Gluons have helicity, essentially a direction of "twist" relative to their motion. In professional jargon, this is called helicity, either positive or negative. Long-standing "folk wisdom" in amplitude theory held that if any number of gluons are involved in a collision, but only one has negative helicity, then the tree-level amplitude must be zero, meaning the process is "forbidden." About a year ago, several theorists noticed a loophole. Zero doesn't always occur when considering extreme configurations where the particles fly in almost the same direction.
Then came the routine, which in theory often proves more difficult than epiphanies. A team led by Andrew Strominger attempted to manually generalize the expressions found from four gluons to five, six, and beyond. The formulas grew, dozens of terms appeared, but a neat, unified notation, as was often the case in similar problems, couldn't be found. Then Alexandru Lupsasca, who was at the time working on enhancing the scientific capabilities of OpenAI models, stepped in and suggested using the problem as a stress test for the model.
Physicists gave GPT-5.2 Pro a cumbersome formula for four gluons and asked it to simplify it. The model completed the task in about 20 minutes, then handled cases with five and six gluons in the same way, reducing long sums to compact products. The key moment came when it was asked to "guess" the general form of the formula for an arbitrary number of particles. The answer arrived quickly, and verification revealed no errors. The general formula was later sent to another, internal OpenAI prototype, which produced a detailed proof, which was also subjected to human verification.
From a scientific perspective, the result sounds neat and not at all like a tabloid sensation. In a preprint on arXiv, the authors demonstrate that "single-minus" amplitudes at the tree level can indeed be non-zero, but not "everywhere," rather in a limited kinematic regime that circumvented the old restrictions. The text includes formulations about half-collinear configurations, as well as about complex continuation of momenta and the so-called Klein space. This doesn't cancel the previous intuition for ordinary, "general" collisions, but it does break the "strictly zero at one minus" thesis as a universal rule.
A second layer of discussion quickly emerged around this story, this time about the role of AI. Physicists quoted by journalists emphasize that the ideas were obvious; the revolution wasn't a new physics "out of nowhere," but rather that a machine was able to push through the thicket of algebra and suggest a compact, general form where humans had been treading water for years. The community's reaction was rather muted: AI can speed up menial tasks, help find errors, and bridge knowledge between fields, but there remains a risk that someone will begin to hide the use of such tools or that the very routine tasks that graduate students typically learn will disappear.
It would seem that a chatbot is a far cry from theoretical physics. But in February 2026, physicists were seriously discussing a result ChatGPT had led them to at an AAAS meeting. It concerned an "impossible" interaction between gluons, the particles that bind quarks into protons and neutrons. For decades, it was believed that such a process simply didn't occur at the most basic level. Now it turns out that it's possible, but it's hidden under very specific conditions, similar to what occurs in the dense and chaotic interior of nucleons.
Gluons mediate the strong force. It's so powerful and nonlinear that even "simple" gluon collisions are described by formulas that quickly give you a headache. Physicists reduce each such collision to a scattering amplitude, a mathematical expression that allows you to calculate the probability of the process. The problem is that these amplitudes grow rapidly, and for just a few particles, it's difficult to simplify them, let alone even accurately finalize them.
Gluons have helicity, essentially a direction of "twist" relative to their motion. In professional jargon, this is called helicity, either positive or negative. Long-standing "folk wisdom" in amplitude theory held that if any number of gluons are involved in a collision, but only one has negative helicity, then the tree-level amplitude must be zero, meaning the process is "forbidden." About a year ago, several theorists noticed a loophole. Zero doesn't always occur when considering extreme configurations where the particles fly in almost the same direction.
Then came the routine, which in theory often proves more difficult than epiphanies. A team led by Andrew Strominger attempted to manually generalize the expressions found from four gluons to five, six, and beyond. The formulas grew, dozens of terms appeared, but a neat, unified notation, as was often the case in similar problems, couldn't be found. Then Alexandru Lupsasca, who was at the time working on enhancing the scientific capabilities of OpenAI models, stepped in and suggested using the problem as a stress test for the model.
Physicists gave GPT-5.2 Pro a cumbersome formula for four gluons and asked it to simplify it. The model completed the task in about 20 minutes, then handled cases with five and six gluons in the same way, reducing long sums to compact products. The key moment came when it was asked to "guess" the general form of the formula for an arbitrary number of particles. The answer arrived quickly, and verification revealed no errors. The general formula was later sent to another, internal OpenAI prototype, which produced a detailed proof, which was also subjected to human verification.
From a scientific perspective, the result sounds neat and not at all like a tabloid sensation. In a preprint on arXiv, the authors demonstrate that "single-minus" amplitudes at the tree level can indeed be non-zero, but not "everywhere," rather in a limited kinematic regime that circumvented the old restrictions. The text includes formulations about half-collinear configurations, as well as about complex continuation of momenta and the so-called Klein space. This doesn't cancel the previous intuition for ordinary, "general" collisions, but it does break the "strictly zero at one minus" thesis as a universal rule.
A second layer of discussion quickly emerged around this story, this time about the role of AI. Physicists quoted by journalists emphasize that the ideas were obvious; the revolution wasn't a new physics "out of nowhere," but rather that a machine was able to push through the thicket of algebra and suggest a compact, general form where humans had been treading water for years. The community's reaction was rather muted: AI can speed up menial tasks, help find errors, and bridge knowledge between fields, but there remains a risk that someone will begin to hide the use of such tools or that the very routine tasks that graduate students typically learn will disappear.
