In an Age of AI, a Physicist Seeks What Endures
Sarah Demers, chair of the physics department at Yale University, sees the potential for large language models to both help and harm her field. The post In an Age of AI, a Physicist Seeks What Endures first appeared on Quanta Magazine.
In July 2026, the U.S. Department of Energy funded 278 projects under its Genesis Mission, which aims to integrate artificial intelligence (AI) into scientific research. One of these projects involves using AI to optimize the Mu2e experiment at Fermi National Accelerator Laboratory in Chicago. The goal is to help physicists search for a rare particle transformation, where a muon (a subatomic particle) becomes an electron. AI can adjust experimental parameters such as magnetic field strength, target placement, and system activation times. Sarah Demers, a physicist at Yale University and chair of its physics department, highlights that AI helps automate time-consuming tuning processes, which can feel more like an art than a precise science. For example, AI can quickly explore combinations of settings to find the most effective configuration, saving researchers significant effort.
Despite AI's benefits, Demers expresses reservations about its use in physics, particularly regarding intellectual property and attribution. Large language models (LLMs) like ChatGPT or Claude are trained on vast datasets that include published works, including those of researchers like Demers, without proper credit. This raises ethical questions about ownership and recognition in scientific work. Demers is leading an effort through the American Physical Society (APS) to create a policy statement on AI use in physics, emphasizing the importance of maintaining skepticism and ensuring that AI does not undermine the integrity of scientific attribution. She argues that physicists must validate AI-generated information to avoid adding noise to the scientific process.
AI is rapidly transforming theoretical physics by accelerating calculations and changing the skills required for problem-solving. Demers notes that some physics Ph.D. programs previously required students to learn German to read important research papers, but AI is now altering these skill requirements. The field is still determining which skills are essential for future physicists. AI tools can help researchers quickly generate drafts of code or data analysis frameworks, allowing them to focus more on interpreting results and formulating new questions. This shift could enable faster progress, but it also demands that physicists adapt their training methods to incorporate these new tools.
On the experimental side, AI is being tested in particle physics to improve data analysis and reconstruction. For instance, the ATLAS experiment at the Large Hadron Collider generates data from hundreds of millions of electronic channels, which must be processed to reconstruct particle interactions. AI can help format this data, create initial data access frameworks, and even revive old experiments by resurrecting outdated computing systems or poorly documented datasets. This allows physicists to revisit historical data and ask new questions that were previously unanswerable. However, Demers cautions that this requires careful validation to ensure the data and methods are reliable.
AI has the potential to change how new physicists are trained. Demers suggests that AI could reduce the time spent on technical tasks like debugging code or accessing data, allowing students to focus earlier on big-picture questions and interdisciplinary research. For example, an undergraduate might spend months learning to create meaningful plots from data, but AI could shorten this process, enabling students to engage with advanced research sooner. However, Demers warns that junior researchers may struggle to use AI effectively without proper guidance, as they might lack the experience to validate AI-generated outputs or recognize when they are going off track.
Experienced physicists may use AI as a thought partner to refine ideas, generate references, and explore unconventional hypotheses. Demers notes that some colleagues use LLMs to quickly test and iterate on research questions, which can be productive for those with deep expertise. However, she advises caution for less experienced researchers, as AI might lead them astray without the necessary background to critically assess its outputs. Demers herself does not use LLMs for writing or research, as she prefers to engage directly with challenges to foster learning and connections. She also worries that AI-generated responses to her co-authored book on physics and dance could reduce opportunities for meaningful conversations with readers.
Demers has never used LLMs like *ChatGPT* or *Claude*, despite her involvement in AI-driven physics projects. As chair of Yale's physics department and a leader in the APS, she feels a responsibility to understand AI's role in the field to advocate for researchers and educate others. While her research group is experimenting with AI agents for the *Genesis Mission*, she does not rely on LLMs for personal tasks such as writing emails or planning. She believes that avoiding AI in these contexts helps her address the root of challenges rather than bypassing them. Demers also emphasizes the importance of human interaction in learning and scientific progress, which AI cannot fully replicate.
