Three Leading Physicists Cite AI in Their Papers, Signaling a Shift in Theoretical Physics Research

Three top physicists consecutively acknowledge AI in their papers, theoretical physics research is undergoing transformation

In the theoretical physics community, Edward Witten is a figure that is hard to overlook. This American theoretical physicist has received one of the highest honors in mathematics, the Fields Medal, and is one of the most influential scholars in string theory, quantum field theory, and mathematical physics. His research has not only profoundly impacted modern physics but also propelled the development of several branches of mathematics.

Now, this giant of physics has begun to use and publicly acknowledge AI in his own papers.

Recently, Weiten uploaded a new paper on arXiv, studying how to define gauge theories with chiral fermions on a lattice. In one technical detail of the paper, he specifically mentioned that Anthropic’s Claude Opus 5.5 suggested a possible mathematical construction. In the acknowledgments at the end of the paper, he also wrote that many details of the research were assisted by Claude Opus 5.5.

This is not the first time Weiten has mentioned AI in a paper. In a paper he published in June of this year, he already disclosed Claude’s assistance in a footnote. Moreover, the physical challenges involved in this research are not being explored by Weiten alone with AI. Within just one month, two other theoretical physicists also published related studies and disclosed the involvement of large models in their papers.

AI proposes a mathematical construction scheme

Weiting’s new paper “A Note On Lattice Definition of Gauge Theories With Chiral Fermions” discusses a problem that has troubled physicists for decades: how to simulate a class of fundamental particles with special properties on a computer.

To study interactions between particles, physicists use a method called “lattice gauge theory.” You can think of it as a grid made of many small squares, where continuous spacetime is divided into tiny cells, and researchers perform calculations using the information on these lattice points. This approach has been applied to problems such as the strong interaction, but it encounters a difficulty when dealing with certain special particles.

For example, some elementary particles possess the “chirality” that physicists speak of, which can be roughly understood as a “left‑handed” or “right‑handed” property. When researchers try to place these particles on a lattice, extra copies of the particles may appear in the calculation, affecting the description of the original physical processes. This is the famous “fermion doubling problem.” In 1981, physicists Nielsen and Ninomiya presented a related theorem that revealed the mathematical constraints involved. In the following decades, researchers have tried many solutions, but establishing a complete chiral gauge theory on the lattice remains a challenging task.

Weinberg discusses a new approach in this paper: first place the gauge fields that describe particle interactions on the lattice, then extend this discrete information to continuous space and combine it with chiral fermions in the continuum for study.In simple terms, it means recording the physical information on a grid first, and then finding a way to fill in the gaps between the grid points so that the entire theory can hold in continuous space.

Image | Edward Witten (Source: Quanta Magazine)

Claude’s suggestion appears right in this process. Witten notes in the paper that an existing method can be insufficiently smooth at the boundaries between adjacent grid cells. He then specifically remarks that Claude Opus 5.5 proposes a possible improvement: by adjusting the coordinate mapping, the gauge field can be smoothly extended into the interior of each grid cell, allowing adjacent cells to connect naturally while ensuring the extended result corresponds precisely to the original information on the grid.

In addition, Claude also proposed an alternative solution that uses a relatively simple linear interpolation method, replacing the originally required step‑by‑step complex mathematical process. Weiting incorporated these suggestions into the main text of the paper and noted in the acknowledgments that Claude Opus 5.5 also helped with many details of the research.

(Source: arXiv)

The paper also notes that, in principle, this method can be implemented on a computer, but it currently faces practical computational difficulties and cannot yet be directly calculated using standard Monte Carlo methods.

In fact, this is not the first time Wei‑Teng has mentioned AI in a paper. In June of this year, he posted a study on “Entanglement Wedge Reconstruction” in holographic theory, discussing the relationship between reconstruction error in quantum gravity and the gravitational geometry. In a footnote of that paper, Wei‑Teng wrote that Claude Opus 4.8 helped him extend a calculation that originally applied only to simple special cases to a more general situation.

Both papers involve AI helping him with concrete mathematical calculations and theoretical constructions. After this paper was released, Weiting mentioned Claude’s content, which also attracted a lot of attention. Columbia University mathematician Peter Woit discussed this in his blog “Not Even Wrong.” He said, “Seeing AI assist outstanding scientists like Weiting in this way to further advance research is encouraging.”

(Source: Peter Woit’s blog)

For the same physics problem, two other researchers also leveraged AI

Weiteng is not the only researcher using AI on this physics problem. Prior to him, two other physicists had already tried to use large models to find solutions, and AI played an even larger role.

On October 6, theoretical physicist Nathaniel Craig uploaded a related paper to arXiv, attempting to solve how to describe chiral fermions in computers. Unlike Wei Ting, who mainly asked AI for a specific mathematical construction idea, Craig involved AI in a broader range of mathematical derivations. In the acknowledgments of the paper he explicitly wrote: “OpenAI’s GPT-6 Astra model was essential to this result.”

(Source: arXiv)

He explains that the research originally stemmed from his own idea, followed by multiple rounds of discussion and revision with GPT‑6 Astra and other experts, gradually shaping the proof presented in the paper. Much of the mathematical derivation was performed by GPT‑6 Astra and then refined and organized by Craig. In other words, the AI not only helped check the calculations but also participated in the process from research concept to concrete proof.

To verify whether these derivations were correct, Craig also had the model use Lean for formal verification. Lean is a mathematical proof‑checking tool that can translate complex derivations into a form that a computer can step‑by‑step verify. The research team designed appropriate tests, which the AI then executed to uncover any potential issues in the proof process.

A few days later, another physicist, Stephen D. H. Hsu, also uploaded a related paper. He adopted a different research approach, attempting to place part of the physical fields on a discrete lattice while retaining a continuous description for the remaining particles, thereby circumventing the difficulties encountered by traditional lattice methods.

In the paper, Hsu disclosed that he extensively used GPT and Claude during his research. Some technical details in the appendix were primarily generated by AI, with GPT handling most of the work, while Claude and other models participated in checking and verification. He also emphasized that he remains responsible for the paper’s results.

On November 6, Hsu further discussed his communication with Weiting on the social platform X. He said that Weiting had shared a draft of this new paper with him, and their research used a similar approach, but the way they employed AI differed. Hsu primarily relied on GPT for the research, while using Claude to check results and look for errors. He also noted that, for this type of research task, he believes GPT’s capabilities surpass Claude’s, and that relatively weaker models can handle verification and error‑checking: if a model finds an issue, it can be passed to a stronger model for judgment.

(Source: social media X)

These papers have also sparked discussion in the physics community and among netizens. Peter Woit dedicated a blog post to analyzing Craig’s research. He believes that having AI help researchers venture into fields they are not originally familiar with is something worth paying attention to. However, he also cautions that this work is still a continuation of existing research directions and remains far from truly solving the underlying theoretical problems. He also noted that he has not examined every complex mathematical detail in the papers.

Another theoretical physicist, Matt Strassler, also discussed this research in his blog. He noted that a large portion of Craig’s paper was generated with the assistance of AI, and although the author has worked to organize the content to be more readable for humans, other researchers still need to invest a great deal of time to truly understand the derivations. For such a lengthy paper dense with technical details, peer review itself is also a substantial undertaking.

(Source: Peter Woit blog)

The physics community on Reddit has also seen related discussions. Some users think this could be a noteworthy development in theoretical physics; others point out that Craig’s paper exceeds 100 pages and raises a practical issue: even if AI can quickly generate massive amounts of mathematical derivations, do human researchers have enough time and energy to check each one? Some also argue that even if an AI‑generated proof appears complete, it still requires expert researchers to understand the underlying logic, and the result cannot be accepted solely on the model’s output.

Currently, the papers by Craig, Hsu, and Weiting are still in preprint stage, and the results need to be vetted by more researchers. However, one thing is already clear: AI is entering the research processes of top scientists.

Reference links:

1.https://arxiv.org/abs/2611.05051

2.https://www.quantamagazine.org/edward-witten-ponders-the-nature-of-reality-20171128/

3.https://doi.org/10.48550/arXiv.2606.18639

4.https://doi.org/10.48550/arXiv.2610.09093

5.https://doi.org/10.48550/arXiv.2610.29019

6.https://www.reddit.com/r/Physics/comments/1x0nz3c/a_gaugeinvariant_measure_for_lattice_chiral

Operations/Formatting: He Chenlong

Note: Cover/hero image generated with AI assistance