AI can write the thesis, but can it develop the scholar?

Generative artificial intelligence is transforming postgraduate research. It helps students engage with unfamiliar literature and improve their writing. Increasingly, it also shapes the reasoning behind their arguments.

These possibilities raise a fundamental educational question: what happens when a thesis develops faster than the person writing it?

A polished chapter may suggest scholarly progress, yet its author may lack the understanding needed to defend its argument. The quality of the text can exceed the researcher's intellectual grasp.

In my recent article (Fataar, 2026), I argue that generative AI is driving an epistemic reconfiguration of higher education. AI is changing how knowledge is produced and validated, while reshaping how students encounter and develop knowledge. These changes raise fundamental questions about the university's educational purpose.

For postgraduate education and thesis work, the challenge is to ensure that AI strengthens disciplinary understanding and independent scholarly judgement rather than bypassing the intellectual work through which these capacities develop.

A thesis develops a scholar, not merely a document

A thesis produces two things simultaneously: a research document and a developing scholar. The document presents an evidence-based argument. The researcher develops the capacity to construct and defend it.

Kamler and Thomson (2007) describe doctoral writing as both text work and identity work. Writing is not merely a technical activity following research. It is part of becoming a scholar.

Through sustained engagement with literature, researchers learn how knowledge is produced within their disciplines. They gradually develop the capacity to position their arguments within scholarly conversations.

This development also has a psychological dimension. Researchers learn to confront uncertainty and respond to criticism. Their confidence grows as they become more capable of explaining their decisions.

AI introduces a tension. It can improve a thesis without necessarily strengthening its author's understanding. For example, a literature review may sound convincing even when the researcher has not engaged deeply with the original studies.

For multilingual researchers, AI can provide valuable linguistic support. Yet polished language should not become the measure of scholarly capability. What matters is whether researchers understand and can defend the knowledge claims they make.

Universities must therefore distinguish between producing a better thesis document and developing a more capable scholar. These outcomes should reinforce each other, but they are not identical.

When AI removes the struggle that develops understanding

AI-assisted research has psychological consequences. An immediate explanation can ease the anxiety associated with grappling with difficult concepts. Yet relief does not necessarily indicate understanding.

Gao and Zhang (2026) examine cognitive offloading and metacognitive calibration in AI-mediated doctoral learning. Their research highlights the danger of students overestimating their understanding when AI provides convincing explanations.

This is known as metacognitive miscalibration, which occurs when students think they understand something better than they actually do. A student may read an AI-generated explanation and feel confident about it, yet struggle to explain its reasoning or show how the evidence supports its conclusion.

Fluent writing can intensify this problem. Coherent language may create an impression of mastery even when the underlying reasoning remains poorly understood.

Cognitive offloading is not inherently harmful. Researchers routinely use technologies to support their work. Such assistance can remove obstacles and create opportunities for deeper engagement.

The crucial distinction is between AI assisting students’ thinking and substituting for it. Assistance occurs when AI helps students engage with difficult material while they remain responsible for understanding and interpreting it. For example, AI might help locate a challenging article, which the student then reads and analyses independently.

Substitution occurs when students accept an AI-generated interpretation without examining the reasoning or evidence behind it. The difference is not simply what AI produces, but whether students use its output to develop their own understanding or allow it to replace their thinking.

Productive intellectual struggle involves confronting uncertainty and evaluating competing explanations. This effort helps researchers develop independent understanding. The aim is not to make learning unnecessarily difficult but to preserve the thinking through which scholarly capability develops.

AI should remove barriers to thinking without removing the thinking itself.

Scholarly judgement cannot be delegated

AI increasingly enters activities central to knowledge production. It can suggest research questions and propose analytical approaches. Yet plausible outputs are not necessarily defensible conclusions.

Scholarly claims require a warrant: a reasoned explanation of why particular evidence supports a conclusion. Researchers must establish this connection according to the standards of their disciplines.

In qualitative research, AI may suggest themes from interviews. Researchers must return to the transcripts and determine whether those interpretations are warranted. In quantitative research, AI may propose a statistical model, but the researcher remains responsible for its assumptions and validity.

AI can also influence scholarly judgement through subtle language changes. Zhao (2025) examines how AI-assisted doctoral writing can reshape authorial stance, including expressions of certainty.

Consider an AI revision that changes ‘the findings may suggest’ to ‘the findings demonstrate’. The revised sentence sounds more authoritative, but the evidence has not changed. Such alterations are not merely stylistic. They change the strength of the knowledge claim.

Scholarly voice emerges through these decisions. It is expressed in how researchers interpret evidence and justify conclusions. Voice is therefore more than writing style: it reflects intellectual judgement and responsibility.

Accountable reconstruction and recapitulation

Universities should neither prohibit AI nor encourage students to depend on it. Instead, they should develop pedagogical practices that help students use AI to strengthen their own thinking and become independent scholars.

Gao and Zhang (2026) propose accountable reconstruction, emphasising the importance of transforming AI-generated material into understanding that researchers can defend. This involves verifying suggestions and taking responsibility for the reasoning behind them.

Building on this approach, I propose recapitulation as a pedagogical practice for postgraduate research. By recapitulation, I mean deliberately returning to AI-assisted work and reconstructing the reasoning behind a claim until the researcher can independently explain and defend it.

Recapitulation goes beyond repetition or paraphrase. It requires researchers to recover intellectual work that AI may have performed on their behalf.

The distinction is significant. Accountable reconstruction emphasises responsibility for knowledge claims. Recapitulation emphasises the learning process through which researchers develop the understanding necessary to exercise that responsibility.

Students could begin by developing their own provisional explanation before consulting AI. They might then invite AI to challenge their interpretation. Crucially, they must return to the original evidence and relevant literature to establish what can be justified.

Close reading remains indispensable. Scholarly texts reveal how arguments are developed and defended. AI-generated summaries may support engagement, but they cannot replace the intellectual development that comes from sustained reading.

From AI interpretation to scholarly understanding

Consider a doctoral student investigating why first-generation university students struggle to develop a sense of belonging. After analysing several interviews, the student asks AI to help interpret the findings. AI produces a convincing explanation suggesting that students' difficulties arise primarily from limited cultural capital.

Recapitulation requires the student to put this polished explanation aside and return to the interview transcripts. Do participants' accounts support this interpretation? Might institutional exclusion provide a more convincing explanation?

The student reconstructs the argument independently, drawing on the original evidence and relevant literature. The revised interpretation may be more cautious, but it is grounded in reasoning the student understands and can defend.

The educational value lies in this return to intellectual work. Rather than merely accepting or correcting AI's interpretation, the student develops independent interpretive judgement. Recapitulation turns an AI-assisted answer into an opportunity for scholarly learning.

This practice should become part of routine thesis supervision. Supervisors should move beyond asking whether AI was used. They should examine what students understand about AI-assisted work and why they accepted particular suggestions.

Universities must also act. Postgraduate development programmes should teach students to evaluate AI-generated explanations and reconstruct arguments independently. Institutional guidance must extend beyond declarations of AI use and concerns about misconduct.

A brief research decision trail could support this process. Students could record significant AI-assisted choices and explain their reasoning. The purpose would be intellectual reflection, not burdensome surveillance.

The responsibility is shared. Students must own their reasoning. Supervisors must make that reasoning central to their conversations. Universities must embed these practices in postgraduate education.

Three questions should become routine in AI-assisted thesis work: Do I understand it? Can I reconstruct it? Can I defend it?

The future of postgraduate education cannot be judged by how quickly AI helps produce completed theses. Its deeper achievement must be the development of scholars capable of contributing credible knowledge. The decisive test is not whether students can produce a convincing thesis with AI, but whether they can stand behind its claims without it.

Aslam Fataar is professor in the Faculty of Education at Stellenbosch University in South Africa. This article is based on his recent webinar, “Research and Thesis Work with AI: Developing scholarly voice, judgement and intellectual ownership”, prepared for master's and doctoral students as part of Stellenbosch University's Postgraduate Office webinar programme.

This article is a commentary. Commentary articles are the opinion of the author and do not necessarily reflect the views of University World News.

References

• Fataar, A (2026). “Epistemic Reconfiguration in Higher Education: Generative artificial intelligence, stratification and the renewal of pedagogical judgement”. Transformation in Higher Education, 11, Article a790.

• Gao, L and Zhang, F (2026). “Cognitive Offloading and Metacognitive Calibration in Generative AI-mediated Doctoral Learning: A grounded theory study of accountable reconstruction”. Frontiers in Psychology, 17, Article 1920947.

• Kamler, B and Thomson, P (2007). “Rethinking doctoral writing as text work and identity work”. In B Somekh and TA Schwandt (Eds.), Knowledge Production: Research work in interesting times (pp 166-179). Routledge.

• Zhao, W (2025). “Reconstructing Stance in EFL Doctoral Thesis Writing through Generative Artificial Intelligence”. Humanities and Social Sciences Communications, 12, Article 1963.