Opinion

How We Rethink the PhD Will Shape the Future of Science

    Verónica Sanz
    • Department of Theoretical Physics, University of Valencia, Valencia, Spain
• Physics 19, 87
AI could make science stronger—but only if we deliberately use it to strengthen the formation of future scientists rather than to reduce their numbers.
APS/Carin Cain
The ability of AI to perform tasks once assigned to PhD students may tempt institutions to reduce the number of doctoral positions. But what would be the long-term impact on science as a whole?

As AI becomes increasingly capable of performing research tasks, an old question acquires new urgency: What is a PhD really for? Not in the bureaucratic sense nor in terms of papers produced or project needs. Rather, I mean something more fundamental: What role can doctoral training play in keeping science alive as a human and intellectual system?

Used well, AI can be genuinely valuable. It can already assist with literature reviews, coding, drafting, brainstorming, and more. But that is precisely why the deeper question matters. If established researchers, working with increasingly powerful AI tools, can do more with fewer junior people around them, what happens to training? And what happens to science itself if fewer people are given the time and space needed to become scientists?

These are not separate questions. A system with fewer entry points is not only less open; it is also one in which fewer intellectual trajectories can develop. By reducing opportunities for young researchers, AI may end up narrowing the space of future discovery.

A PhD is not just a stage of productive work. Nor is it simply a credential. It is one of the main ways in which science reproduces and renews itself. Through doctoral training, researchers learn not only techniques but also judgment: how to tell whether a problem is worth pursuing, whether a result is robust, and whether an idea is genuinely interesting or merely elegant looking. They learn how to deal with uncertainty, recover from mistakes, and recognize the difference between a promising direction and a dead end.

That kind of evolution does not come from consuming polished outputs. It emerges from practice, trial and error, and the gradual development of scientific judgment. If too much of that process is bypassed because AI can generate competent intermediate products, we may temporarily preserve productivity but also impair the formation of future scientists.

An immediate concern is access to science. If institutions decide that fewer junior positions are needed because AI makes senior researchers more productive, entry into academic research will narrow. From a short-term efficiency perspective, that may look rational: fewer positions; similar or greater output. With scarcer opportunities, entry into science would depend more heavily on conventional markers of promise: educational background, social networks, institutional prestige, and access to opportunities early in one’s career. The system becomes more closed and less able to recognize talent that does not immediately appear in familiar forms.

But the loss doesn’t only affect those excluded: Science itself loses possibilities, as fewer people form independent judgments or pursue questions that initially look marginal.

It is also important to not overstate AI’s discovery power. It can identify patterns, synthesize fields, suggest hypotheses, and propose apparently promising directions. But we should not confuse plausibility with insight, or fluency with originality. Science does not advance only by extending what already looks promising. Very often, the most important advances emerge from unexpected or marginal directions rather than obviously promising ones. Quantum mechanics did not arise from a defined program aimed at building a new theory of nature. It emerged from tackling problems such as explaining blackbody radiation and other anomalies that did not fit the existing picture.

This is where too much reliance on AI becomes risky. Systems trained on existing literatures and prevailing definitions of relevance may direct many researchers toward the same questions, arguments, and candidate explanations. What looks like acceleration may also be a form of convergence: science becoming better at developing conventional ideas and worse at noticing what does not fit.

Younger researchers are not inherently more original than established ones, and it would be naive to romanticize inexperience. But a system that trains many people, through different paths, creates more possibilities for independent thinking and unexpected directions to emerge. Science needs distributed exploration, not only concentrated optimization.

None of this means that students should be shielded from useful tools or required—in the name of training—to continue performing repetitive tasks that AI can do more efficiently. AI could make doctoral education better, leaving more time for understanding, exploration, and independent thinking. But that will happen only if we are deliberate about using it to enrich the formation of young scientists rather than to reduce the number of people allowed to become scientists at all.

So the real question is not whether AI can do more of the work that junior researchers once did. Clearly, it can. The question is what we think those junior researchers were there for. If they were mainly a source of labor, shifting some of their tasks to AI may look like an obvious gain. But if doctoral training is one of the main ways that science creates new perspectives and possibilities for discovery, then shrinking those pathways is a decision about the future of science itself.

That is why the future of doctoral training in the age of AI is not a secondary issue. We need strong junior cohorts in AI-assisted environments, students trained to use AI critically rather than passively, and evaluation systems able to recognize originality and intellectual courage when fluent output is easier to produce. Otherwise, we may discover too late that in making science faster, we have also made it narrower.

About the Author

Image of Verónica Sanz

Verónica Sanz is a professor of theoretical physics at the University of Valencia in Spain and a researcher at the Institute for Corpuscular Physic (IFIC)—a joint research center of the University of Valencia and Spanish National Research Council (CSIC). Her research focuses on particle physics beyond the standard model, effective field theories, axions, cosmology, and the use of machine learning and artificial intelligence in fundamental physics. She has held positions at universities in Europe and North America and has been involved in several major international collaborations. Sanz is also active in knowledge transfer and interdisciplinary applications of AI.


Recent Articles

Winding Spins Toward Better Quantum Sensing
Magnetism

Winding Spins Toward Better Quantum Sensing

Robust collective spin dynamics have been achieved in diamond by winding the spins into a nanoscale spiral and making their interactions symmetric. Read More »

Quantum Dots Emit Photons in Desired Directions
Photonics

Quantum Dots Emit Photons in Desired Directions

Researchers demonstrate directional control of a single-photon source by coupling two quantum dots together. Read More »

The Surprising Origin of Uranus’ Moons and Rings
Astrophysics

The Surprising Origin of Uranus’ Moons and Rings

New infrared observations tie the ice giant’s small moons and narrow rings to a past encounter with a visiting object from the Kuiper belt. Read More »

More Articles