Supervising PhD Researchers in the Age of AI

At the start of a PhD, asking a supervisor ‘What should I do next?’ is perfectly reasonable. By the end, I hope the student is more likely to arrive with a view of what is worth investigating next, and why.

That change is not always visible in a progress report. Experiments may become more sophisticated, presentations more polished and papers more numerous, while the important decisions about the research still come from someone else. Generative AI can make the distinction harder to see. It helps students read unfamiliar material, write code, organise papers and communicate clearly. I welcome much of that assistance. But completing tasks efficiently is different from learning to decide which tasks matter.

To me, doctoral supervision is partly about helping a student make that transition: from contributing to a research project to exercising informed judgement about its direction.

Figure 1. From AI-assisted routine tasks to independent scientific judgement. Conceptual illustration generated using AI; people and research graphics are illustrative.

A research question is not always the student's to begin with

Many doctoral projects begin with a question framed by the supervisor. The funding may already be in place; the equipment, collaborators and broad methodology may be fixed. A new student cannot reasonably be expected to challenge every assumption before they understand the literature or the experimental system.

At this stage, technical guidance matters. Students need to know how to use the instruments, what the established methods can tell them, and what constitutes a fair comparison. They also need clear expectations and feedback they can act on. Independence does not begin by pretending that expertise is unnecessary.

Yet implementing a plan should not become the whole doctorate. As students gain experience, I want them to notice when the evidence no longer fits the original question. Perhaps the planned experiment cannot distinguish two explanations. Perhaps an unexpected observation is more interesting than the result the project was designed to obtain. Perhaps the method is working, but the claim made for it is too broad.

Formulating a useful research question takes practice. Reading more papers helps, but reading alone does not teach someone how to identify the uncertainty that matters. That develops through repeated choices: what to measure, what to compare, what to rule out and when to reconsider an assumption.

When a better result does not answer a better question

Robotics offers many opportunities to improve a performance metric without learning why a system behaves as it does. A higher grasp success rate or lower perception error can be an excellent engineering achievement. It does not automatically reveal which part of the system made the difference, or whether the improvement will hold under different conditions.

The problem becomes especially interesting across physical scales. For a conventional robot arm, adjusting the controller may compensate for some disturbances. In microrobotics, geometry, available actuation and interactions with the surrounding fluid or nearby surfaces may place different limits on motion. Which effects dominate depends on the particular robot and operating conditions; the word ‘microscale’ is not itself an explanation.

So I would not always begin by asking which controller performs best. I would want to know whether the difficulty comes from control, from an inadequate model or from a physical limitation that control alone cannot overcome. If the robot can change configuration, we can also ask whether that change makes a previously infeasible task possible. Each explanation calls for different evidence.

An experiment that cannot distinguish between these possibilities may yield another improvement in the benchmark while leaving the scientific question untouched. Students should learn to recognise that point. Sometimes the right decision is to calibrate the system more carefully, vary one physical condition or narrow the claim before developing a more complex learning method.

A tightly scoped technical contribution is still worthwhile. I do not expect every PhD experiment to uncover a general principle. I do expect the researcher to understand what the result demonstrates, and what it does not.

What should happen in a supervision meeting?

A progress meeting can easily become a tour of completed tasks: code written, experiments run, figures produced. Those details are useful, but I would rather spend time on the choices behind them. Why this experiment? What other explanations were considered? What would we conclude if the result went in the opposite direction?

A supervisor's role is not simply to supply the next technical answer. Experience may help us see an overlooked assumption, connect findings from different fields or recognise an experiment that is unlikely to settle the dispute. The more valuable response may be a question that leads the student to reconsider the problem.

I also want students to disagree with me when they have a reason. A proposed direction may turn out to be impractical, or I may have missed something important. Either way, examining the evidence is more useful than treating the supervisor's first suggestion as an instruction that cannot be challenged. My experience is relevant, but it does not make every judgement correct.

This matters in a multidisciplinary research group. Someone working on tactile perception may interpret a result differently from a colleague focused on control or physical modelling. The point is not to make everyone an expert in every discipline. It is to help students recognise when another perspective changes the question, rather than merely adding another technique.

Independence needs more supervision than silence

There is no single moment when a doctoral student becomes independent. A student may be confident in one part of a project and need close guidance in another. An experienced researcher learning a new experimental method may need the same detailed support as someone at the beginning of their PhD.

What can change gradually is who proposes the options. Early meetings may focus on helping a student design a workable experiment. Later, the student might bring several approaches, explain the trade-offs and recommend one. Eventually, I would hope they can identify an important gap, decide how to investigate it and revise their plans when the results demand it.

That progression depends on reliable support: regular discussion, clear expectations, access to equipment and expertise, and timely, specific feedback. Students should be able to bring an inconclusive result or an unfinished idea to a meeting without feeling obliged to disguise it as progress.

It is tempting to equate autonomy with leaving students alone. I think that is a mistake. The challenge for a supervisor is to remain engaged without making every scientific decision on the student's behalf.

Learning from experiments that do not go to plan

Physical experiments rarely follow a tidy sequence. Measurements drift, calibrations fail and models leave out effects that matter. An unexpected result may point to a genuine limitation, or it may turn out to be an artefact. Neither interpretation should be assumed at the outset.

Students need experience deciding what to check next. Would repeating the measurement help? Is there a calibration or a control experiment that separates the explanations? Does the result undermine the hypothesis, or only the way it was tested? Sometimes the most sensible choice is to seek advice from someone with different expertise.

I am also cautious about the way universities sometimes discuss resilience. Persistence is useful, but it should not mean tolerating unlimited pressure or repeatedly pursuing a failing plan. Students are more likely to examine uncertainty openly when supervision makes room for honest discussion, realistic workloads and decisions to change course. Good research training should make it possible to report a failure without having to present it as a personal failing.

AI can assist the work; it cannot own the judgement

AI can be especially helpful when a student is entering a new field, debugging routine code or improving academic English. I would not judge their research negatively simply because they used it for those purposes. Nor should we insist that every line of code or every explanation be produced unaided.

The more revealing questions are about decisions. Why was a method chosen? What assumptions does it make? Was an AI-generated reference checked against the original paper? If the tool offered several solutions, which ones did the student test, and why was one accepted?

A useful disclosure of AI assistance can explain those decisions rather than merely listing the tools used. For instance, a student can distinguish code generated to implement a specified procedure from an analytical choice made after comparing alternatives. The first may be a sensible delegation. The second needs scientific justification, regardless of who proposed it.

Time saved also presents a choice. If AI makes routine work quicker, students might read the source papers more carefully, repeat a difficult experiment or explore a result they did not expect. Or the saved time may simply be filled with more slides, more plots and a longer task list. Supervisors influence that choice through what they ask about and what they reward.

What I would look for in a PhD progress review

Publications, technical milestones and presentations matter. They show what has been achieved and communicated. But a review should also reveal whether the student is becoming better at directing research.

I would want to hear how they chose an experiment, what the evidence supports, which assumptions remain uncertain and how the next step follows from those findings. I would also ask whether the research question has changed, and what led them to change it.

This need not become an oral examination designed to catch students out. An uncertain answer that is carefully reasoned tells a supervisor something useful. A fluent but rehearsed answer may tell us less. For students working in a second language, the discussion should make room for their scientific reasoning rather than judging it mainly by how quickly they can express it in English.

Originality also needs a standard. A proposal does not become good merely because it differs from the supervisor's idea. It must be informed by the literature, meaningful in relation to the evidence and possible to investigate. Learning to make that judgement is a central part of the doctorate.

By the end of a PhD, I hope a student can recognise a worthwhile problem, explain why it remains unresolved and investigate it with appropriate methods. AI may help them do that work. The supervisor's lasting contribution is to help them develop the curiosity, confidence and critical judgement to decide what the work should be.

AI assistance was used to edit the language and structure of this article. The arguments and views are the author's own.