PhD students are increasingly using generative AI (LLMs) to write, revise and communicate their research. For supervisors, this raises a difficult question: how can they recognise what students genuinely understand when AI can produce convincing academic text? Yet AI may also create opportunities for deeper scientific discussions and more independent thinking. This article explores how AI is changing trust, intellectual ownership and assessment in PhD supervision, and which questions supervisors and students should be asking each other.
1. When good writing no longer tells the whole story
Imagine receiving a manuscript from your PhD student. The argument appears coherent, the structure is convincing, and the language is considerably better than in previous drafts. You are pleased with the progress, but you also wonder how much of the text was produced with AI. Did the student develop the argument independently and use AI to improve the wording? Did AI help them organise their thinking? Or did they accept suggestions without examining them sufficiently?
All options can produce remarkably similar texts, yet they represent very different levels of intellectual engagement. For years, supervisors have used students’ writing to understand how they think and where they are in their project progress. An unclear introduction might reveal uncertainty about the research problem, while a weak discussion could indicate difficulties interpreting the results. Generative AI is making these signals less reliable because it can improve the presentation without necessarily improving the underlying reasoning. The text may sound fluent, the thinking may not live up to the level of fluency.
In our academic writing courses, researchers increasingly tell us that AI has become part of their everyday writing process. Some supervisors welcome the opportunities, while others worry that students may lose essential skills or become dependent on tools they cannot critically evaluate. Both reactions are understandable since fluent academic language can conceal unsupported claims, or questionable interpretations.
The challenge for supervision is to identify whether a polished manuscript is a result of excellent scientific thinking, effective AI assistance or a combination of both. It may also conceal weaknesses that would previously have been easier to recognise.
For supervisors, the question is becoming: How can we understand the thinking behind a manuscript when the writing itself no longer tells us the whole story?
2. Does AI replace thinking or make deeper thinking possible?
A concern we frequently hear, particularly from established researchers, is that researchers who use AI will gradually stop thinking independently. If a student asks AI to formulate an introduction, summarise literature, develop a discussion, or suggest a good paper title, they may avoid the intellectual effort that would otherwise help them develop essential research skills. Academic writing has always been closely connected to thinking, and the difficulty of explaining an argument is an important part of learning.
Using AI does not necessarily mean someone has done less intellectual work. Consider a student who asks AI to suggest alternative interpretations of a finding, compares those suggestions with own evidence, rejects unsupported explanations and develops a stronger argument. That student may engage in more critical thinking than someone who writes independently but never seriously questions their initial interpretation.
The educational value depends on how the student engages with the technology. AI can provide convenient answers that are accepted without scrutiny, but it can also expose weaknesses and encourage researchers to consider alternatives. The difficulty is that supervisors cannot reliably distinguish these processes simply by reading the final text. The fluency of a linguistically well-crafted text created by an LLM can perfectly disguise the intellectual process behind manuscript.
Thus, doctoral education faces a real dilemma: Students need to develop the ability to construct scientific arguments, but they also need to learn how to evaluate arguments critically, including those generated by AI. Both abilities matter, and their relative importance may change as researchers become more experienced.
A question to answer for supervisors and students: Which intellectual skills should be developed through independent practice, and where could AI help deepen the learning process?
3. Trust, ownership and the changing relationship
PhD supervision depends on trust. Supervisors need confidence that students understand their research and take responsibility for their work. Students need confidence that their contributions will be recognised and that their supervisors will assess them fairly. AI can complicate both sides of this relationship, particularly when expectations about its use remain unspoken.
A supervisor may become suspicious when a student’s writing suddenly improves, even when AI was used only to clarify carefully developed ideas. Conversely, a student may avoid discussing AI use because they are uncertain how their supervisor will respond. Undisclosed use of AI tools may lead to an unspoken lack of trust in the supervisor-student relationship. The willingness to openly discuss assumptions and expectations may decrease.
AI also raises questions about ownership. Researchers have always developed ideas through conversations with colleagues, feedback from supervisors and engagement with literature. AI adds another source of suggestions, although one that can produce convincing but unreliable arguments. The important question is whether students critically examine these suggestions and take responsibility for the scientific decisions that follow.
Recently, a participant in one of our writing courses described a difficult experience. After submitting a manuscript, the student’s supervisor said that the paper was so poorly written that it was difficult to understand. Rather than discussing the problems with the student, the supervisor rewrote almost the entire manuscript and returned the revised version with little explanation beyond saying that this was how the paper should be written.
The student was unhappy with the result. Several changes seemed scientifically questionable, and some appeared factually incorrect. Yet without an explanation of the supervisor’s reasoning, it was difficult to understand what had been changed and why. The student considered using AI to compare the two versions, identify the substantive differences and examine whether the revised arguments were justified but the fear of a negative judgement prevented this step.
This example raises questions that extend well beyond AI. Of course supervisors may have the skill to rewrite a manuscript and produce a better text. Yet, if this happens in a closed environment, what can the student learn from that process? Has it helped them understand how to improve their own writing or become a more independent researcher? And what happens to trust when a student’s work is replaced rather than discussed?
The situation also illustrates an interesting opportunity for AI. It could help the student examine the changes more systematically, identify potential problems and prepare questions for a conversation with the supervisor. AI cannot determine who is scientifically correct, and its suggestions would still require critical evaluation. But it might help the student engage more confidently with feedback from someone in a position of greater academic authority.
The example reminds us that intellectual independence involves more than producing original ideas. It also requires the ability to question arguments, defend a critical decisions and even engage critically with the contributions of others, including supervisors. Ideally, supervision should help students develop precisely these abilities.
A question to reflect for supervisors and students: Does the way you exchange feedback strengthen intellectual independence and trust, or could it unintentionally undermine both?
4. Could AI make supervision more intellectually demanding?
If written work reveals less about a student’s thinking, supervisors may need to change how they engage with manuscripts. A familiar process involves students submitting drafts, supervisors adding comments and students revising accordingly. This can be effective, but much of the interaction concentrates on the written product rather than the reasoning behind it.
AI makes the limitations of this approach more apparent. Instead of focusing primarily on whether the text is well written, supervisors may need to ask students to explain why they selected a particular interpretation, which alternatives they considered and how the evidence supports their conclusions. Such conversations provide a better understanding of the student’s intellectual development and may also reveal weaknesses that are difficult to identify from the manuscript alone.
We can also consider it as an opportunity to make supervision more productive. Imagine a student using AI to explore several explanations for an unexpected result before a meeting. The supervisor and student could then examine these possibilities together, compare them with the evidence and identify where additional analysis is needed. The student would need to defend their decisions, while the supervisor could challenge assumptions and contribute scientific expertise.
The way how supervision time is spent could change: If AI helps students improve language and structure, meetings may devote less attention to correcting text and more to developing arguments and discussing scientific decisions. These are activities central to doctoral education, and they offer opportunities to strengthen critical thinking and independence.
None of this comes automatically just because AI sits on the table as well. Students need sufficient disciplinary knowledge to recognise weak or misleading AI suggestions. Less experienced researchers may still need to practise developing arguments without assistance before they can evaluate alternatives confidently. Supervisors therefore need to consider how AI use relates to the student’s stage of development.
A question for supervisors: Do your supervision meetings mainly improve the manuscript, or do they also help you understand and develop the researcher’s scientific judgement?
5. Are we still assessing the right skills?
The implications extend beyond individual supervision meetings. A PhD is an educational process through which researchers develop the ability to conduct independent scientific work. Traditionally, supervisors have assessed this development through research activities, discussions, presentations and written outputs. AI makes the relationship between completing a task and demonstrating competence less straightforward.
A well-written literature review does not necessarily demonstrate that a student has critically examined the literature. A convincing discussion does not guarantee that they can defend the interpretation. These limitations existed before AI, but the technology makes them more visible because polished outputs can now be produced with less direct intellectual engagement and at faster speed.
Supervisors and doctoral programmes may have to pay greater attention to how students arrive at their conclusions. Asking students to explain methodological choices, compare alternative interpretations and defend scientific decisions could provide more meaningful evidence of their developing competence. Written work remains essential, but it may need to be complemented by more direct engagement with the reasoning behind it.
For doctoral programmes and institutions: Do current supervision and assessment practices still provide a reliable picture of students’ independent research competence?
6. The conversation we need to have
Generative AI is changing how PhD students research, write and communicate. It creates genuine risks when students become dependent on suggestions they cannot evaluate, when fluent writing conceals weak reasoning or when uncertainty about AI use undermines trust. It may, however, also create opportunities for deeper scientific discussions, stronger critical evaluation and more independent thinking.
The outcome will depend partly on how supervisors and students respond. Rules about acceptable AI use, disclosure and research integrity remain necessary, but they cannot resolve every educational question. Two students may use the same tool in very different ways, with very different consequences for their learning. Understanding those differences requires more than knowing whether AI was used.
We believe the most useful starting point is an open conversation about expectations and intellectual development. Supervisors and students need to discuss which skills should be developed independently, where AI can provide valuable support, how scientific decisions should be justified and what transparency means in their working relationship.
For institutions, the challenge is broader. Doctoral programmes need to consider whether their approaches to supervision and assessment still reflect the abilities they want researchers to develop. They also need to support supervisors who may themselves be uncertain about how AI changes the educational relationship.
For years, supervisors have relied heavily on written work to understand how their PhD students think. AI means that the finished text may no longer reveal that thinking as clearly as it once did. Perhaps this is an opportunity to make supervision more interactive and more focused on the intellectual decisions behind the research.
The purpose of doctoral education remains the development of researchers who can think critically, make independent scientific decisions and take responsibility for their contributions. AI does not change that purpose, but it may change how we achieve it.
The question worth discussing together is whether our use of AI helps PhD students become more independent researchers, and how we can recognise when it does.
Resources & Further Reading
From the Smart Academics Blog:
- #160: Research Communication in the Age of AI: Why Academic Writing Still Matters — Why fluent text production is not the same as meaningful research communication.
- #157: Why Good PhD Supervision Feels So Difficult — Understanding the challenges of developing independent researchers.
- #159: Why Academic Writing Problems Are Often Institutional Support Problems — Why publication progress also depends on the institutional support environment.
Related Courses
These questions are also relevant to two upcoming TRESS ACADEMIC open courses.
Responsible Use of AI for Academic Authors | 2 December 2026
A live-online workshop exploring how researchers can use generative AI responsibly in academic writing while maintaining intellectual ownership, scientific judgement and research integrity.
How to Improve PhD Supervision | 10–11 December 2026
A live-online course for PhD supervisors who want to reflect on their supervision practices, strengthen communication with doctoral students and support the development of independent researchers.
© 2026 Tress Academic