Reflections on agentic AI, scientific education, and the future of human expertise.
Imagine a young astrophysicist asking an AI assistant to analyse a stellar catalogue. The assistant selects the data, writes the code, produces a beautiful Hertzsprung–Russell diagram, identifies stellar populations, and drafts a paragraph for a research paper. What used to take days might now take minutes.
The student presents the result to a supervisor. The figure looks convincing. Then comes a simple question: Why does this feature appear in the diagram?
What if the student cannot answer without asking the AI again?
This question captures a tension that I think deserves far more attention. Artificial intelligence promises to accelerate science enormously. But could it also weaken the process through which people become scientists?
From doing research to describing research
Computational science has always involved a chain of reasoning that goes well beyond producing a figure. A researcher chooses a dataset, understands its limitations, develops an analysis, writes or adapts software, evaluates uncertainties, and compares results against physical expectations.
In astrophysics, that often means learning Python, C++, or Fortran; debugging a pipeline; understanding numerical techniques; and running analyses or simulations on high-performance computers. These tasks can be tedious. They are also educational.
Today we are beginning to replace parts of that process with a conversation. A researcher describes the desired outcome, and an AI agent plans the steps, writes code, calls tools, accesses data, submits jobs, and presents results. With increasingly capable agentic systems, several stages may run with little human intervention.
Consider a request: "Use Gaia data to investigate the kinematics of stellar populations in the Galactic disc and halo." An AI assistant could retrieve catalogues, apply selections, calculate velocities, create plots, and propose an interpretation. A similar system could configure a cosmological simulation, submit it to a computing cluster, monitor execution, and analyse its output.
I work at the intersection of astrophysics, high-performance computing, and AI-driven workflows. I see the enormous opportunity here. Scientists could spend less time on repetitive technical work and more time exploring ambitious questions. Small research groups could gain access to analyses that previously required specialised teams.
Yet a deeper question remains: If the machine performs the steps, how does the researcher learn why those steps are necessary?
The paradox of expertise
We do not become experts simply by receiving correct answers. We become experts through attempts, mistakes, corrections, and the gradual development of intuition about what can go wrong.
A student who has struggled with a numerical simulation may recognise an instability that a newcomer overlooks. Someone who built a catalogue-processing pipeline may notice a selection effect before trusting an unexpected correlation. A researcher who has repeatedly compared models with observations knows that a beautiful plot can conceal a serious methodological problem.
What happens when an AI removes most of that struggle?
In the short term, students become more productive. They can complete assignments faster, explore more ideas, and create polished results. But if they always delegate the difficult parts, they may never acquire the understanding those difficulties once produced.
The problem is not that students use AI. The problem is that they may learn to trust AI before they learn to challenge it.
Imagine a student presenting an AI-generated colour–magnitude diagram. The supervisor asks how extinction was handled, which stars were excluded, why a particular cut was chosen, or how uncertainties affect the interpretation. The student turns back to the AI for every answer.
At that point, are we educating a scientist or training a person to repeat an assistant?
The uncomfortable word I used in our discussion was parrot. A scientific parrot can repeat a plausible explanation without understanding the reasoning that makes it credible. It may even speak more confidently than a genuinely careful researcher.
Now extend the scenario across several generations of students. Today's experienced scientists often recognise mistakes because they learned their disciplines by doing the underlying work themselves. If tomorrow's students never develop the same foundations, who will train the generation after them?
Taken to its extreme, this is not only a problem for university education. It could threaten the self-correcting capacity of science itself.
Can one AI really check another?
A natural answer is to use AI to verify AI. One agent generates code; another reviews it; a third checks the scientific interpretation. This may indeed improve reliability. Automated tests, independent calculations, benchmarking, and reproducible execution are valuable.
But agreement between several AI systems is not the same as independent evidence. They may share assumptions, blind spots, or systematic errors. An impressive consensus can still be built on a flawed premise.
Suppose a cosmological simulation produces an unexpected structure. Is it a new physical insight? Or an artefact of numerical resolution, initial conditions, boundary treatment, or a subgrid prescription? Another AI may help investigate. It may also provide a confident but incorrect explanation.
The real test must connect the claim to independent observations, robust methods, meaningful uncertainty estimates, and predictions that could prove it wrong.
Human experts are fallible too. That is why "keep a human in the loop" is not a complete answer. We need a competent human in the loop: someone who understands enough to disagree with the machine, design a test, and know when the evidence is insufficient.
A golden era of papers—or of discoveries?
AI could dramatically reduce the effort required to generate scientific manuscripts. It can already help with literature searches, code, visualisation, and writing. Increasingly sophisticated systems could assemble entire analyses and draft papers.
But the number of papers is not a reliable measure of our understanding of nature.
In cosmology and fundamental physics, a model can generate intriguing arguments about dark matter, modified gravity, galaxy formation, or fundamental constants. The hard part is not inventing a hypothesis. The hard part is showing what distinguishes it from alternatives and what observation could falsify it.
If producing scientific claims becomes cheap while validating them remains costly, we could enter an era of exceptionally polished but poorly tested research.
A manuscript is not a discovery. A plausible physical story is not a measurement. And an attractive figure is not, by itself, evidence.
Acceleration is valuable only if the quality of scientific reasoning keeps pace.
Must the next generation still learn to code?
This leads to a question that is more difficult than it first appears: if AI can write scientific software, is it still necessary for scientists to learn programming?
I do not think the answer is a simple yes or no.
Programming has never been the ultimate purpose of astronomy. It is a means of turning questions into tests. Many useful technologies have allowed researchers to work at a higher level of abstraction. We do not require every astronomer to manufacture a detector or implement every numerical algorithm from scratch.
Nevertheless, understanding an analysis is different from merely requesting it.
A researcher must know what a calculation assumes, how uncertainty propagates, how a numerical result can fail, and what alternative explanation remains possible. Practical knowledge of code and computation is often one of the most effective ways to develop that understanding.
Perhaps future scientists will write fewer lines of code. But they must still be able to inspect the logic of an analysis, question the choices made on their behalf, and reconstruct essential results when necessary.
Otherwise we will not have replaced labour with intelligence. We will have replaced understanding with dependency.
Education should change—but not surrender
Banning AI from scientific education is not a serious long-term solution. These tools are becoming part of research, and students should learn how to use them well. Yet giving students unlimited assistance without changing how we assess understanding is equally inadequate.
I would favour a two-level approach.
First, students need protected opportunities to acquire foundations: mathematics, physics, statistics, computational reasoning, and the habit of working through difficult problems. They should sometimes derive a result, diagnose a broken analysis, or defend a conclusion without an AI assistant. The point is not nostalgia for manual work; it is to build the intellectual muscles that make independent judgment possible.
Second, once that competence is developing, students should learn to work with agents critically. They should ask an AI to propose methods, then identify the assumptions; compare independent implementations; create falsification tests; reproduce results; and explain where the machine might be wrong.
A powerful examination question may no longer be "Can you generate this plot?" It may be: "Here is a plot and an explanation produced by an AI. What would you check before believing either of them?"
We should reward the ability to discover an error, not merely the ability to produce an answer.
What kind of scientist do we want to create?
I am optimistic about the possibilities of AI-assisted research. I work on systems intended to make advanced data analysis and scientific computing more accessible. My concern comes from taking that promise seriously, not from wanting to stop technological progress.
AI could give researchers unprecedented freedom to explore new questions. It could improve reproducibility, reduce technical barriers, and allow humans to spend more time on conceptual work.
But intellectual independence is not automatically preserved when tools become more capable. It must be cultivated deliberately.
The future scientist may be less of a programmer and more of a designer of questions, tests, and interpretations. That could be an extraordinary development—provided we do not confuse fluency in directing an AI with expertise in the science being done.
So the challenge is not simply to build AI that can perform research.
It is to build a research culture that continues to produce people capable of understanding, criticising, and improving the research AI performs.
If we succeed, AI may help create a new golden age of discovery. If we fail, we may produce generations of researchers who can ask machines for answers but no longer know how to tell whether those answers are true.
These reflections grew out of an ongoing conversation about AI, research automation, and the education of future scientists. They are intended to provoke discussion, not to predict an inevitable outcome.
Preparation note: This essay developed from my discussion of AI and scientific training, with ChatGPT assistance in structuring and drafting the text. It presents reflections and questions, rather than results from a completed educational study.