Physicists See AI as Tool, Not Replacement, for Research
Physicists attending the Global Physics Summit this year expressed cautious optimism about artificial intelligence, viewing it primarily as a collaborator that speeds up routine tasks while leaving the harder intellectual work to humans.
The consensus emerged from interviews with researchers across multiple disciplines. A postdoc at Virginia Tech said AI can compress a four- or five-hour coding task into a 10- to 15-minute chatbot conversation. Four other attendees reported regularly using generative AI to write code.
Beyond coding, physicists are deploying AI for data analysis and pattern recognition. Orit Peleg, an associate professor at the University of Colorado Boulder who studies animal communication, said AI's ability to identify and classify specific behaviors "opened up a whole new field of quantitative animal behavior" that wasn't feasible before.
Simona Mei, a professor at Paris Cité University working on galaxy classification, presented results showing AI reduced contamination detection errors by 70%, freeing researchers to focus on analysis rather than refining detection code.
Speed of Improvement Surprises Researchers
Nearly every physicist interviewed cited the rapid pace of AI advancement as their biggest surprise. Hilary Egan, a data scientist at the National Laboratory of the Rockies, noted that foundation models now work effectively on complex scientific data with minimal customization-a shift from the past, when developing a single AI model for one specific dataset was standard practice.
Janelle Shane, a laser scientist, said she was struck by how much AI-generated text has improved in coherence over the past decade. However, she cautioned that modern models are less transparent than earlier versions, making them "less of a blank slate and more of a sticker book."
Consensus-Seeking Creates Blind Spots
Several researchers flagged a fundamental limitation: AI tends toward consensus answers, which could stifle the unconventional thinking that drives scientific breakthroughs.
Theodore MacMillan, a PhD candidate at Stanford studying AI weather models, said chatbots struggle to generate genuinely new ideas and can be "sycophantic" about labeling ideas as novel. This makes them unreliable for understanding what work already exists in a field.
Simona Mei expressed concern that AI's tendency toward consensus could hamper "critical thinking, creativity, and originality in research." Matthew Ginsberg from Google told summit attendees: "You all are the best physicists you can be exactly when you are not giving the consensus answer. That is something that large-language models currently are completely incapable of doing."
Egan added another practical constraint: AI needs rapid feedback loops and massive data volumes to learn from mistakes-a challenge when training models to predict outcomes of complex experiments.
Human Judgment Remains Essential
The physicists emphasized that AI requires human oversight. Rachel Burley, APS Chief Publications Officer, said that while AI can make researchers and editors more productive, "having a human in the loop, and AI output always backed by scientific rigor, is critically important."
Matthew Schwartz, a professor at Harvard, identified problem selection as the core skill AI hasn't mastered. "Once you specify the problem clearly, the solution will be almost automatic," he said. "Problem selection is really the fundamental thing that we can do as humans."
Sarah Demers, chair of APS's Panel on Public Affairs, noted that the scientific process-generating hypotheses, conducting experiments, validating results, quantifying uncertainties-doesn't change with AI adoption. Instead, physicists should apply "the same rigor and skepticism that has brought us to our current understanding of the natural world."
Abhijit Chakraborty drew a parallel to programming languages. Just as Python and C++ became tools that scientists are responsible for using correctly, he said, "like any other tool, there has to be a human counterpart to make it accountable."
Domain Expertise Will Differentiate
Shane predicted that domain expertise will become increasingly valuable. Physicists with deep technical knowledge can see past marketing claims and identify "where the AI-shaped problem actually is."
Several interviewees described themselves as "cautiously optimistic" about AI's future in physics. MacMillan captured the sentiment: "There are a lot of things we've got to figure out, like the beginning of any truly great scientific endeavor."
For researchers looking to develop proficiency with these tools, AI Research Courses and AI Data Analysis Courses offer structured training in how to apply AI effectively to scientific work.
Your membership also unlocks:
