Physicist Sarah Demers on AI's Role in Particle Physics

The U.S. Department of Energy selected 278 projects in July 2026 for its Genesis Mission, which aims to integrate artificial intelligence into science. One project will use AI to help physicists search for an extremely rare particle transformation-a muon turning into an electron-by optimizing the Mu2e experiment at Fermilab.

“We have all of these knobs that we can turn to optimize our experiment,” said Sarah Demers, a physicist working on Mu2e and chair of the physics department at Yale University. She told Quanta magazine that AI could help decide factors like magnetic field duration and strength, target placement, and which systems to activate. “Having AI capabilities help us populate that space is a godsend,” she said.

However, Demers holds strong reservations about AI in other areas. She objects to large language models (LLMs) being trained on datasets that include others' intellectual property without proper credit, likely encompassing her own writing. In response to AI's rise, she is leading an effort through the American Physical Society to draft an AI policy statement. She argues physicists must now confront what is enduring about their discipline.

AI in Theory and Experiment

Demers noted the fastest changes are currently in theoretical physics, where new LLMs are massively accelerating calculations. This shift is altering the necessary skill sets for solving problems. She compared it to past requirements for physics Ph.D. Students to learn German to read key papers. The field must now determine which skills remain important and which do not.

On the experimental side, physicists are still exploring AI's potential. For instance, at the ATLAS experiment at the Large Hadron Collider, data from hundreds of millions of electronic channels must be reconstructed to understand particle interactions. AI coding assistants can help draft data access frameworks, potentially speeding up analysis. Another exciting use is revisiting old experiments. Researchers can now ask new questions of archived data, even if it is not perfectly formatted, enabling inquiries that were previously impractical.

Policy, Property, and Attribution

A primary concern for Demers and many peers is intellectual property and attribution. “This is a moral hazard. We can't let this stand,” she stated. She described the tension between outrage at uncompensated use of work and the excitement of having a vast repository of knowledge at one's fingertips. “It's so wrong, and it's so incredible, simultaneously,” she said.

Despite disagreements, Demers said a key area of consensus is that a physicist must be responsible for the content of any publication they sign. Researchers may not need to understand every detail of an AI model, but they must have a way to validate its output. Without this, she warned, the field risks adding noise rather than making scientific progress.

Training the Next Generation

Demers envisions a best-case scenario where AI allows new undergraduates to engage with new data much faster. Previously, a new student might take five months to produce a meaningful plot. With AI handling technical hurdles like code bugs and data access, trainees could focus earlier on formulating interesting questions and learning how to answer them.

She contrasted this with her own graduate experience, where admitting ignorance to a colleague was a humbling but valuable learning process that forged human and intellectual connections. For junior researchers today, LLMs offer a judgment-free way to ask questions, though Demers cautions that reaping benefits requires expertise. She noted that more advanced researchers sometimes use LLMs as “thought partners” to quickly get references and refine ideas, but only if they know enough to recognize when the model goes off track.

Demers remains skeptical of hype and insists human physics is not over. The foundational practice of giving correct attribution and maintaining a conversation among contributors must be preserved. Otherwise, she argues, the field risks losing track of what is valid.