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

Synthesis: Saxena et al. report a qualitative study at the University of Bristol: 15 postgraduate research (PGR) students, five per faculty and eight international, on how they use, resist and justify generative AI in doctoral research. Four themes emerged: negotiated acceptable use; Learner Agency, authenticity and resistance; GenAI as or refused 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 use unethical, a conflict read through foreign language anxiety and attribution error. Coding was accepted far more readily than writing, and uploading transcripts was rejected on Privacy and trust grounds. Students called delegation "offloading" or "cheating" and feared losing the process of becoming an independent researcher. Supervisors were enthusiasts, prohibitors or silencers who never raised the topic, pushing some students to GenAI for Feedback they were not getting.

Key Findings

1. Fifteen PGRs at one UK Russell Group university. Five per faculty, stages from MRes to fourth year, seven home to eight international, from around 3,500 doctoral students.
2. Four themes, not one attitude. Negotiating acceptable use; agency, authenticity and resistance; GenAI as surrogate supervisor; and navigating university guidance. After hallucinations, use was "measured, cautious, and accompanied by continuous monitoring".
3. Writing support split the cohort along language lines. International PGRs said GenAI eased expression and confidence; many home students called the same use inappropriate or unethical.
4. Coding was the least contested use. Almost all appraised it positively, and novices built tools without prior programming experience through vibe coding.
5. Data analysis was the most contested use. Uploading a transcript would "betray the trust of my participants"; quantitative data drew fewer objections.
6. Efficiency, framed as a supervised "research assistant". Time-poor participants used GenAI for routine work, ideation and slides, while many treated it as a "glorified search engine".
7. Agency and conscience, not detection, governed behavior. Delegation was "offloading" or "cheating"; one wanted to "not lose connection to my own work".

How the study was designed and analyzed

GenAI literacy research across higher education has grown quickly, but postgraduate research remains thinly covered and existing studies measure use, not judgment. Three questions follow: how PGRs use GenAI, what they see as opportunities and challenges, and how supervisory contexts shape practice. The literature review distinguishes GenAI as tool, assistant and thinking partner. Recruitment used posters, newsletters and email lists plus an expression-of-interest questionnaire. Transcripts were analyzed following Braun and Clarke's reflexive thematic analysis.

Negotiating acceptable use across disciplines and languages

Acceptability, not frequency, is the organizing puzzle, and the sharpest line fell between writing and coding. Writing carried a moral charge because academic writing is treated as evidence of independent identity, so home students without the same linguistic demands projected their standards onto peers, a pattern the authors call fundamental attribution error; foreign language anxiety fit international students better. One participant declined literature searching because paywalled sources were unreachable. Humanities and Social Sciences students reported lower use, citing ethical and environmental reasons.

Agency, authenticity and the imposter cross-reading

Participants feared diminishing human capability: some prized research's messy middle, so delegation risked surrendering growth and satisfaction. Internal standards, not detection risk, governed behavior: one wanted to fix his own code rather than lose connection to his work. The authors read this through the impostor phenomenon: students may fear being "found out" inside accepted boundaries, an intersection with self-doubt they flag for further study. Resistance was unstable: normalization pushed some to conceal their use, while others proposed GenAI as a thinking partner preserving Learner Agency and critical thinking.

Supervisors as enthusiasts, prohibitors and silencers

Where supervisors modeled experimentation, students used tools openly (role models); where they were skeptical, students concealed use, while some assumed supervisors trusted them to keep academic integrity. Reliance grew where support was thin: distant contact and reluctance to ask questions for fear of appearing incompetent, the same Help-Seeking inhibition GenAI removes. Many saw supervisors as less informed than themselves and wanted staff training; chatbots filled the surrogate role. The authors call the silencer stance most harmful, since expecting PGRs to train supervisors adds cognitive burden.

Institutional guidance, recommendations and limitations

Recommendations respond to unclear institutional positions, inequity for international students, overreliance and the impostor phenomenon, and supervisor preparedness. Proposals include core principles that outlast tool generations, explicit supervisory conversations, guidance that is clear and discipline-specific, co-production with PGRs, and embedded training. On equity, participants rejected language barriers being framed as "an international student problem". Assessment is unfinished business: with fluent prose easy to produce, the thesis weakens as a signal and the viva strengthens, though the viva is not proof against GenAI. They ask how thesis, viva, portfolio and practice-based models can respond without losing rigour.

What this means for practice

• Instructors. Raise GenAI explicitly in supervisory conversations: the authors name the "silencer" stance most harmful, and students hid their use where supervisors were skeptical.
• Instructors. Provide language and writing support for all PGRs, rather than letting it become an international-student workaround.
• Administrators. Co-produce guidance with PGRs and make it discipline-specific and broader than writing; participants said it covered maybe 20-30% of research.
• Administrators. Train supervisors and embed that training across the doctoral program, since several students were their supervisor's primary GenAI source.

Limitations

• Fifteen PGRs were purposively sampled from one UK Russell Group university, five per faculty, so findings support transferable claims rather than generalization.
• Supervisors were not interviewed, so the enthusiast/prohibitor/silencer typology rests on students' accounts and cannot be checked.
• Recruitment let participants self-select, plausibly those with developed views, and the paper is a pre-print.
• PGR study is heterogeneous, so a cross-faculty sample of 15 loses interdisciplinary nuance the authors say needs comparative work.

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.