Three Top Physicists Credit AI in Theoretical Physics Papers

Edward Witten, a leading mathematical physicist and Fields Medalist, has publicly credited Anthropic’s Claude Opus 5.5 in a new arXiv paper on defining gauge theories with chiral fermions on a lattice, signaling a growing shift in theoretical physics toward AI-assisted research. The paper addresses the long-standing Nielsen-Ninomiya fermion-doubling problem, a constraint that makes it difficult to simulate chiral fermions using conventional lattice methods. Witten explores a framework in which gauge fields are discretized on a lattice and then extended into continuum space so that chiral fermions can be incorporated consistently. In the paper’s technical discussion, Witten says Claude Opus 5.5 proposed a coordinate-mapping adjustment to smooth gauge-field extensions between adjacent lattice cells and preserve correspondence with the original discrete data. The model also suggested a linear-interpolation alternative to a more complicated stepwise construction. Witten acknowledged Claude in both the main text and acknowledgments, and noted that the method is conceptually computable but not yet practical with standard Monte Carlo techniques. This follows a June paper by Witten on entanglement wedge reconstruction in holographic theory, in which a footnote credited Claude Opus 4.8 with helping generalize a calculation. Witten’s disclosures have drawn attention from commentators such as Peter Woit, who called AI assistance to a top scientist encouraging, while cautioning that progress remains incremental and dependent on human scrutiny. The same lattice-chiral-fermion problem has recently attracted other AI-assisted attempts. On Oct. 6, Nathaniel Craig posted a paper stating that OpenAI’s GPT-6 Astra was essential to the result. Craig said he supplied the initial idea, then iterated with the model and other experts, with GPT-6 Astra generating many mathematical derivations that he revised and organized. The team used Lean formal proof checks to test portions of the derivation. Days later, Stephen D. H. Hsu released a separate paper using a hybrid lattice-continuum approach and disclosed extensive use of GPT and Claude, with GPT performing much of the work in appendix details and Claude and other models assisting in verification. Hsu emphasized that he remained responsible for the findings. He later said Witten had shared his draft and that the researchers used similar approaches with different AI workflows, adding that he found GPT stronger than Claude for generative reasoning, while weaker models could help identify errors for stronger models to evaluate. The papers remain preprints and have not yet undergone full peer review. Discussion in physics blogs and online communities has focused on both the promise of entering complex mathematical territory and the burden of checking lengthy, AI-assisted proofs, including Craig’s more than 100-page manuscript. The broader takeaway is that large language models are moving beyond literature review and basic coding into substantive roles in theoretical derivation, proof exploration, and verification, while human authors still need to assess correctness and interpretability.