From Research Assistant to Surrogate Supervisor: A Qualitative Study Exploring PGR Students' Diverse Uses of Generative AI

Synthesis: Saxena, Tierney, Grist, O'Leary and Palmer report a qualitative study at the University of Bristol based on semi-structured interviews with 15 postgraduate research (PGR) students — five per faculty, eight of them international — about how they use, resist and justify generative AI in doctoral research. Reflexive thematic analysis generated four themes: acceptable use negotiated task by task; Learner Agency, authenticity and resistance; GenAI as, or refusing GenAI as, a surrogate supervisor; and navigating university guidance. International students leaned on GenAI for writing and described linguistic relief, while many home students judged that same use inappropriate or unethical — an inter-student conflict the authors read through foreign language anxiety and fundamental attribution error. Coding support was accepted far more readily than writing support, and uploading qualitative transcripts was rejected outright on Privacy and participant-trust grounds. Students described delegating work as "offloading" or "cheating", invoked conscience, and feared losing the process that makes an independent researcher. Supervisors appeared as enthusiasts, prohibitors or — most troubling to the authors — silencers who never raised the topic, pushing some students towards GenAI for Feedback their supervisors withheld. Guidance was called narrow, writing-focused and shallow; students asked to co-produce policy and argued the viva may matter more as the thesis becomes easier to fake.

Key Findings

1. Fifteen PGRs, purposively spread across faculties and stages. Five participants per faculty across three faculties of one UK Russell Group university, drawn from a doctoral population of around 3,500 (over 1,500 international), with study stages running from MRes (1) through first (4), second (4), third (3) and fourth year (3) and student status split seven home to eight international. Methods are covered below; ethics approval was code 28858.
2. Four themes, not one attitude. (a) Negotiating Acceptable Use of GenAI; (b) Agency, Authenticity, and Resistance; (c) (Resisting) GenAI as Surrogate Supervisor; and (d) Navigating GenAI Use with the University. The authors stress that use was "measured, cautious, and accompanied by continuous monitoring" after participants' encounters with "hallucinations" and "confabulations" in AI output.
3. Writing support split the cohort along language lines. For all international PGRs interviewed, English was likely not their first language, and GenAI reduced the barrier: "Since English isn't my mother tongue, it gives me more confidence to communicate my ideas clearly and have more impactful and accessible writing; it's an aspect I don't feel confident in" (P06). Many home students regarded that use as inappropriate — in P01's account of his peers, "if you don't know how to write, you should not be awarded a PhD because they don't struggle with it" — and a few called it unethical.
4. Coding was the least contested use. Almost all participants appraised GenAI positively for writing code, troubleshooting errors and learning specialist software, and novices acquired skills without prior programming experience: "The way of being effective is changing [with] AI. Now everyone can code even without having any IT [knowledge]. I can spend a couple of hours researching… and, [through] vibe coding, train my own AI app" (P02).
5. Data analysis was the most contested use. Only a small number reported using GenAI for analysis, and several objected strongly to uploading data: "I find it really problematic to upload, for instance, a transcript just to summarise or select some quotes. [Using GenAI] would betray the trust of my participants [and] will undermine their privacy" (P06). Participants handling quantitative data saw fewer risks, treating numbers as relatively decontextualised, while qualitative transcripts were read as contextual and sensitive.
6. Efficiency, framed as a supervised "research assistant". Time-poor participants used GenAI to skip routine and administrative work, plus ideation, slides, abstracts and job applications: "If you use it in a responsible way… as a research assistant beside you [that] helps you think, write, [conduct] research… [then] it can speed up the research significantly" (P11). Many others treated it as a "glorified search engine", checking outputs rather than accepting them "at face value", and one avoided literature searching altogether because paywalled sources were inaccessible.
7. Agency, authenticity and conscience. Delegating intellectual work was described as "offloading" and as akin to "cheating", with the value of research located in process as well as product: "I want to know that I can do these tasks for myself… I think I don't want to lose connection to my own work" (P07). Even use that could not be detected carried a cost: "even if the world doesn't know, my conscience will know, and it won't sit right with me" (P04). The authors connect this to the impostor phenomenon in doctoral study.
8. Supervisors ranged from enthusiasts to prohibitors to silencers. Some participants reported little or no discussion of GenAI use; others called it a "taboo" topic both sides avoided. A student whose supervisor was an "AI researcher" experimented more freely, while P01 concealed use: "I don't discuss this [GenAI use] with my supervisor… based on casual conversations… he's very sceptical about use of [GenAI]". Where supervisory support felt distant or judgmental, GenAI became a surrogate supervisor, and some students resisted the delegation: "you might not know the answer, but I'm coming to you as the source of information and not GenAI" (P09).
9. Guidance was confusing, narrow and shallow. "I think it's very tricky to make them [guidance] clear. There's a very hard line to walk" (P07); for an international PGR, "it's very conservative and narrow… It's only about writing… [which] is maybe 20-30% of research" (P11). Students wanted case studies, accessible language, discipline-specific guidance, training embedded across the degree rather than one induction, and co-production: "this has to be a very community-grounded approach. I think students have to act as voices within that decision-making process" (P15). Several proposed greater weight on the viva: "You can't fake them because you have to give them yourself" (P12).

How the study was designed and analysed

The paper starts from a gap rather than a controversy: GenAI literacy research across higher education has expanded quickly, but postgraduate research remains thinly covered, and where studies exist they measure use rather than how researchers judge it. Three research questions follow — how PGRs use GenAI in research, what they see as its opportunities, challenges and appropriate use, and how supervisory and institutional contexts shape their practice. The literature review frames the analysis with a distinction between GenAI as tool (bounded tasks), assistant (explanation and feedback) and thinking partner (participation in developing and interrogating ideas), noting that these roles allocate responsibility differently and that chatbots give more immediate, task-focused feedback than supervisors do (Jensen et al., 2026). It also names the paradox the findings later explore: GenAI can raise a student's functional autonomy while displacing the dialogic work through which epistemic independence develops.

Recruitment ran through physical posters, PGR newsletters and faculty email lists, followed by an expression-of-interest questionnaire; purposive selection then balanced faculty, discipline, year and home or international status. Semi-structured interviews were conducted by the lead author in person or online, with the interview guide shared in advance so participants could reflect first, and participants were encouraged to use non-identifying language and offered a £10 voucher. Auto-generated Teams transcripts were checked, anonymised and pseudonymised, then exported to NVivo. Analysis was inductive, grounded in a contextualist epistemology that treats knowledge as shaped by the researcher's positionality, and followed Braun and Clarke's (2021) six phases of reflexive thematic analysis. All five authors wrote positionality statements — a final-year international PGR, a senior lecturer and PGR supervisor, an academic staff developer focused on AI, the head of the institute, and a senior education developer — and the group reflected collectively on how their backgrounds and concerns about GenAI might shape interpretation.

The paper also situates institutional guidance comparatively: the University of York emphasises critical oversight, intellectual property and researcher development; King's College London permits defined uses in thesis writing subject to declaration and supervisory discussion; and the University of Bristol permits proofreading and diagnostic feedback while forbidding substantial thesis generation, recently asking students to confirm they have read its guidance and that the dissertation complies with it.

Negotiating acceptable use across disciplines and languages

Acceptability, not frequency, is the study's organising puzzle: some students used GenAI often, others "sparingly" or not at all, for personal, social, ethical, moral, intellectual and environmental reasons, and every reported use sat inside ongoing negotiation and compromise. The dividing line the authors press hardest is between writing and coding. Writing support carried a moral charge because academic writing is treated as evidence of an independent academic identity, so home students who did not face the same linguistic demands projected their own standards onto peers who did — a pattern the authors explain with Ross's (1977) fundamental attribution error, where others' behaviour is read as character and one's own as circumstance. For international students the situation is better described by foreign language anxiety (Horwitz et al., 1986): discomfort, nervousness and fear of negative evaluation when working in a non-native language, which P01 connects to dreading judgement over an email or a group message. Coding, by contrast, was recognised by both groups as a legitimate barrier deserving support, even where coding was central to the degree.

Trust in the technology was conditional throughout. Awareness of inaccuracies made credibility a live question, and reviewing AI output was standard practice, particularly for literature work, where one participant noted that paywalled evidence is unreachable and so declined to search that way at all. Usage also varied by discipline, with students from Humanities and Social Sciences reporting lower use and stronger resistance, and it varied by task beyond research, extending to presentations, illustrations, conference abstracts and job applications.

Agency, authenticity and the imposter cross-reading

The second theme asks what GenAI does to a researcher's sense of ownership. Participants worried about diminishing human capability and about the future role of people in knowledge generation; some described the messy, uncertain middle of research as the site of pleasure and skill development, so that delegation — "offloading" — risked surrendering both intellectual growth and satisfaction. P07's account of wanting to fix his own code and "not lose connection to my own work" and P04's conscience quote are the paper's clearest statements that internal standards, not detection risk, governed behaviour. The authors read these through the impostor phenomenon (Clance and Imes, 1978), arguing that students may fear being "found out" even when their use sits inside institutionally accepted boundaries, and that this novel intersection with self-doubt deserves further study.

Resistance, however, was not stable. Participants observed the normalisation of GenAI in academic and professional life and worried about competitiveness: "The scepticism is still there but… [GenAI] is here to stay… People will be using this in any profession, [and] you might be behind if you're not" (P09). Some used GenAI strategically and concealed it, describing anxiety, guilt and shame at the prospect of being judged, while others proposed a more sustainable position — GenAI as a "thinking partner" that supports Creativity and efficiency while preserving Learner Agency, authenticity and critical thinking.

Supervisors as enthusiasts, prohibitors and silencers

Supervision is where the study's most transferable claims sit. Whether GenAI was discussed at all depended on the relationship: some students assumed their supervisors trusted them to maintain academic integrity and saw no need for the conversation, while others described deliberate avoidance on both sides. Where supervisors modelled experimentation, students used tools more openly and talked about them freely, consistent with supervisors as role models; where supervisors were sceptical, students hid their use. Reliance on GenAI grew where support was thin — distant, hierarchical or infrequent supervisory contact, and reluctance to ask questions for fear of appearing incompetent, the same Help-Seeking inhibition that GenAI removes. Participants also felt the transition to independence was rushed without sufficient guidance, and GenAI filled the gap as a surrogate supervisor offering immediate, personalised, iterative feedback at any hour.

A second finding is asymmetry of expertise. Many students saw their supervisors as less informed than they were about GenAI and called for formal staff training; in some cases students became their supervisors' primary source of information, which participants themselves flagged as risky, since a student's interpretation of appropriate use may be wrong and an under-informed supervisor cannot sense-check it. Others rejected the surrogate role outright, insisting that supervisors should hold the expertise to guide doctoral work rather than redirect students to a chatbot. The authors name three supervisor stances — enthusiast, prohibitor and silencer — and single out the silencer as the most harmful, arguing supervisors have a duty of care to model and discuss responsible use, and that expecting PGRs to train their supervisors convolutes power dynamics and adds cognitive burden.

Institutional guidance, recommendations and limitations

The final theme concerns what universities can do, and the study's recommendations respond to six stated concerns: limited clarity about institutional positions, limited guidance, inter-PGR conflict and inequities for international students, GenAI avoidance and rejection, overreliance and the impostor phenomenon, and supervisor preparedness. The proposals are specific — core research principles that outlast each tool generation; explicit supervisory conversations about boundaries; guidance that is clear, discipline-specific, accessible, signposted and extends beyond writing to ethics and research use; co-production with PGRs; and training embedded across the programme rather than delivered once. The authors also insist that students who decline GenAI are neither advantaged nor disadvantaged, and that institutions provide language and academic writing support for all PGRs. On equity, participants argued that boundaries should be clear for everyone while some students are permitted to navigate closer to them, and one framed this as institutional "reciprocal accountability" for structural disadvantage, rejecting the view that language barriers are simply "an international student problem". Assessment, finally, is treated as unfinished business: with GenAI able to produce fluent prose, participants suggested the thesis becomes a weaker signal of scholarly ability and the viva a stronger one — a claim the authors accept only partly, since the viva is not inherently proof against GenAI and the real question is how doctoral assessment can evidence agency, judgement and ownership.

The limits are stated plainly. Fifteen students at one UK university, purposively sampled, cannot ground generalisation; PGR study is heterogeneous in expectations, supervision and GenAI acceptance, so interdisciplinary nuance is lost; supervisors were not interviewed, so their perspective is inferred from students; and the work is a pre-print. The authors ask for longitudinal research tracking confidence, researcher identity and the impostor phenomenon over time, for comparative work across disciplines and student groups, and for designs that include supervisors alongside students, together with research on how thesis, viva, portfolio and practice-based models can respond to quickly changing tools without giving up rigour or integrity.

Connected Concepts

• Generative AI — the technology PGRs negotiated task by task across the doctorate
• Academic Integrity — the frame students used to judge writing help, and the site of inter-student disagreement
• Learner Agency — researcher agency, authenticity and the refusal to delegate the intellectual process
• Cognitive Offloading — "offloading" read as cheating, plus the process and skill development students feared losing
• AI Literacy — the literacy students said institutions had not built, and that they partly built for their supervisors
• AI Use and Disclosure Statements — concealment, guilt and shame, against the case for openly declared GenAI use
• Hallucination Risk — hallucinations and "confabulations" as the reason verification became continuous
• Privacy — qualitative transcripts as too sensitive to upload, against quantitative data seen as decontextualised
• Multilingual Learning — foreign language anxiety and GenAI as writing support for non-native English speakers
• Equity — international students, structural disadvantage and institutional reciprocal accountability
• Learner Identity — researcher identity, intellectual ownership and the impostor phenomenon cross-reading
• Teacher AI Competency — supervisors as enthusiasts, prohibitors or silencers whose own GenAI knowledge was uneven

Connected Articles

• Academic integrity in the age of generative AI: A scoping review of research on higher education student voices — Scoping review of student voices on academic integrity with generative AI
• Tools facilitate cheating, or partner supports learning? GenAI and academic integrity issues from students' perspectives — Competing student readings of GenAI as cheating tool or learning partner
• Is using artificial intelligence tools for academic work cheating? Student perceptions, ethics, and the impact — Student perceptions of whether AI use for academic work counts as cheating
• "Should I Tell My Teacher?" Student AI Disclosure Practices, Stigma, and Self-Regulated Learning in Higher Education — Disclosure, stigma and self-regulated learning in student GenAI use
• To disclose or not to disclose: Peer influence and psychological factors in students' use of generative artificial intelligence — Peer influence and psychological factors in disclosing generative AI use
• Student Use of and Views on GenAI for Writing — Student use of and views on GenAI for academic writing
• Understanding Student Dependency on AI: The Role of AI Literacy, Academic Self-Efficacy, and Resource Management Strategies — Dependency on AI considered alongside literacy and academic self-efficacy
• Understanding student perspectives on generative AI chatbots: a human-centred mixed-methods study in higher education — Human-centred mixed-methods account of student perspectives on AI chatbots
• Governing generative AI in higher education: a global Delphi study on policy and practice — Global Delphi study on institutional policy and practice for generative AI
• Are Students Dependent on AI in Writing Courses? Analyzing Factors Influencing Dependence on Generative AI Through the I-PACE Model — Factors behind dependence on generative AI in writing courses

Citation

Saxena, G., Tierney, A., Grist, H., O'Leary, R., & Palmer, A. (2026). From Research Assistant to Surrogate Supervisor: A Qualitative Study Exploring PGR Students' Diverse Uses of Generative AI. University of Bristol. Pre-print, Version 1, 24 August 2026.