ChatGPT helped derive a formula for the interaction of elementary massless particles.
For 40 years, describing how gluons (elementary massless particles) interact with each other was considered an extremely complex and almost impossible to calculate problem. Recently, physicists incorporated GPT-5.2 Pro into the solution and, in several steps, derived a generalized formula.
The scientists verified the formula and confirmed a scientific breakthrough: doubts about LLM hallucinations were disproved. Interestingly, OpenAI's internal model, codenamed "SuperChat," which is still under development, participated in the verification.
https://www.science.org/content/article/chatgpt-spits-out-surprising-insight-particle-physics
in case someone wonders why this hasn't been attacked much by humans:
Gluons are "force carrier" virtual particles; you're probably familiar with photons, so gluons are the same thing, except they have no existence outside of the nucleus in any regular situation.
Understanding them better might help interpret crazy high-energy cosmic rays, or improve the understanding of why each nucleus has its binding energy, although they have zero effect on things that only deal with electromagnetism, and the edge cases investigated in this research don't even come up in standard nuclear physics.
It's still a cool use of models in a rigorous domain, and I hope it's not simply the same kind of models with more training data and parameters... because that would still not be 100% reliable, as opposed to something which incorporates some fundamentally new mechanism for the rigorous reasoning, and could be as reliable as a theorem-proving system.