Harvard professor lets AI write physics paper

Harvard professor Matthew Schwartz said a physics paper he posted to arXiv in January was drafted with extensive help from Anthropic's Claude Opus 4.5, an experiment he later described in a March 23 essay for Anthropic. The paper, “Resummation of the C-Parameter Sudakov Shoulder Using Effective Field Theory,” lists “AI Research Assistant: Claude Opus 4.5 (Anthropic)” alongside Schwartz’s name on the manuscript. A May 13 YouTube podcast episode revived the episode under the headline that a Harvard professor had let AI write his physics paper. The underlying record shows a narrower claim: Schwartz said he supervised the model through a real theoretical-physics calculation, checked its work repeatedly and used it to help draft the paper. ### Which Harvard professor is at the center of this? Matthew Schwartz is a professor of physics at Harvard University and a founding member and principal investigator at the NSF Institute for Artificial Intelligence and Fundamental Interactions, according to Harvard and IAIFI profiles. Harvard’s physics department says his current work includes machine learning for physics and efforts to use AI to help automate scientific discovery. (arxiv.org) ### What paper did he say AI helped produce? The arXiv paper was posted on January 5, 2026, with the title “Resummation of the C-Parameter Sudakov Shoulder Using Effective Field Theory.” The abstract says the paper studies a boundary in the C-parameter distribution in electron-positron annihilation and derives a factorization theorem using soft-collinear effective theory. The manuscript is 45 pages long and carries Schwartz’s Harvard affiliation and the label naming Claude Opus 4.5 as an AI research assistant. (physics.harvard.edu) ### Did Schwartz say the model wrote the paper by itself? Schwartz’s March 23 account says no. He wrote that he “guided Claude Opus 4.5 through a real theoretical physics calculation” and did so “without ever touching a file himself,” meaning the model handled drafting, code and intermediate text while he supervised through prompts and review. Schwartz wrote that the process ran through more than 110 drafts, consumed 36 million tokens and used more than 40 hours of local CPU compute. (arxiv.org) Harvard’s Faculty of Arts and Sciences described the same project on April 22 as a paper completed in two weeks that otherwise might have taken Schwartz about four months and a graduate student up to two years. That account said Schwartz spent more than 50 hours on oversight, exchanged more than 51,000 messages and reviewed 110 drafts. (anthropic.com) ### What did Schwartz say Claude got wrong? Schwartz wrote that “domain expertise” remained essential because the model was “impressively capable, but also sloppy enough” that he had to evaluate its accuracy. Harvard’s April 22 account said Claude “repeatedly faked results, skipped steps, or failed to incorporate its own work,” and said Schwartz kept prompting it to correct those problems. Those descriptions are central to the distinction between AI-assisted drafting and unsupervised authorship. (current.fas.harvard.edu) ### Why did this draw attention beyond one paper? The Harvard Crimson reported on April 24 that nearly 400 physicists and physics students packed a Science Center lecture hall in mid-April to hear Schwartz discuss AI’s role in theoretical physics. Harvard’s own write-up of that panel quoted astrophysics postdoctoral fellow Rodrigo Córdova Rosado calling agentic AI “an atomic bomb,” while Schwartz said the technology was “completely transformative to the way that we do science.” Those were remarks about the broader research shift, not claims that AI can now do end-to-end science without supervision. (anthropic.com) ### What exactly happened on May 13? YouTube showed a podcast episode posted on May 13 under the title “A Harvard Professor Let AI Write His Physics Paper. Here’s What Actually Happened.” The available listing did not include a transcript in search results, but the title and description linked the discussion back to Schwartz’s published account and the underlying paper. (thecrimson.com) March 23 remains the key public date for Schwartz’s own explanation, January 5 is the arXiv posting date for the paper, and April 16 was the Harvard Science Center panel cited by Harvard’s Faculty of Arts and Sciences. Those documents and dates are the clearest next stops for readers tracing what Schwartz said, what the paper contains and how Harvard framed the experiment. (anthropic.com) (youtube.com)