AI Is Now Designing Physics Experiments. The Discipline Is Arguing About What's Left for Humans.

In a Tübingen lab, an artificial intelligence rearranges lasers, mirrors, and lenses to design an experiment. At Fermi National Accelerator Laboratory, banks of fine-tuning controls on the Mu2e experiment, a search for a muon spontaneously transforming into an electron, retune themselves while the beam is running. Both are now standard practice, and a new Nature review by Klimesch, Arlt, Ruiz-Gonzalez, and colleagues argues the pattern is broader: experimental physics is becoming a search problem over hardware-configuration space.

The review, published in Nature 657, 47-58 (2026), organizes the field around four research questions: how to build expressive search spaces; how to make fast, reliable simulators of the experiment; how to translate a scientific goal into a computable objective function; and how to navigate both discrete and continuous design choices. Applications already span quantum optics, electron microscopy, fusion devices, particle detectors, and gravitational-wave observatories, according to institutional press materials from TU Wien and the Tübingen AI Center. The review's central claim is that AI-discovered configurations "often challenge established design conventions while matching or even exceeding the performance of human-designed set-ups."

Mario Krenn, who leads the AI experimental-design work at the Tübingen AI Center and the university's Cluster of Excellence "New Perspectives for Science," has been the public face of this program since his group's early quantum optics work produced an experimental setup the team had not been able to devise themselves. The Nature review is in part an attempt to systematize the trick Krenn's group has been pulling: define the configuration space broadly enough, score it with a fast surrogate model, and let an algorithm walk it.

The same approach is now migrating into the routine part of running an experiment, not just designing it. At Fermilab, the Mu2e experiment's parameter control, the dozens of fine adjustments that used to require hand-tuning during beam runs, is being turned over to automated systems, a shift that Phys.org's scientific-editor coverage places inside the same move from manual calibration toward AI-driven search. The transition is subtle: the optics and magnets are not being redesigned, but the path to a stable, high-rate configuration is now being optimized rather than handed down.

That is the part of the story the physics community is now arguing over. Sarah Demers, chair of physics at Yale and a co-author of an upcoming American Physical Society policy statement on AI usage, frames the shift as a discipline-level renegotiation, according to HotHardware's coverage of her remarks. Routine calculation, experimental design, and device optimization are moving to algorithms, in her account, while the durable human skill set she names is narrower and more specific: scientific skepticism, proper attribution of intellectual breakthroughs, and the safeguarding of human-to-human critical discourse.

The three concerns are not interchangeable. Skepticism is the practice of challenging a result on grounds the algorithm cannot supply. Attribution is the question of who gets credit when an AI proposes the configuration that produces the paper-worthy measurement. Critical discourse is the laboratory and conference-floor conversation where one physicist pushes another on a result, in person, until the claim either survives or collapses. If those three erode, the search can keep producing experiments while the discipline loses the apparatus that decides which experiments matter.

The Nature review and Phys.org's coverage both note that large language models and specialized software suites are accelerating symbolic derivations, compressing months of algebraic work into days. The implication is not that theorists disappear, but that the analytical skill set expected of a new doctoral student is shifting underneath them, and graduate programs have not yet settled on a replacement.

Demers's upcoming policy statement is one attempt to write down the answer to a question the field has not settled: whether physics will train the next generation in the human-specific work, the skepticism, the attribution norms, the face-to-face argument, at the same rate it is handing the rest of the craft to algorithms. The statement will land before the practice ossifies around tools that did not ask permission.