AI evidence extraction

Strategies for AI Use in Thesis Writing among Undergraduate Language Students: A Think-Aloud Protocol

Ramiaida Darmi, Nurkhamimi Zainuddin, Noor Saazai Mat Saad, Fariza Puteh-Behak · Asia Pacific Journal of Educators and Education · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

1/4

Quality (LMQS)

Evidence

Citations

0.00

FWCI

This is an AI analysis. Read the peer-reviewed original at the publisher: https://doi.org/10.21315/apjee2026.41.1.18

Methodology & findings

Study design

Qualitative research design employing Retrospective Think-Aloud (RTA) protocol.

Sample

N = 8, 5 groups

Primary method

Qualitative thematic analysis following Braun and Clarke's (2006) six-phase procedure. No quantitative statistical tests or confidence intervals were reported. Data managed using ATLAS.ti software. Analysis involved iterative manual coding, theme development using inductive grouping, and interpretive scaffolding from AI Literacy Framework (OECD, 2025) and PAIRR model (Sperber et al., 2025). Data saturation was assessed, with the authors noting "Data saturation was reached by the sixth interview; however, two additional participants were included to confirm the recurrence of key patterns and ensure analytical comprehensiveness."

Main result

The study found that students "use it like everywhere, from the beginning until the very end" (Fina), demonstrating comprehensive AI utilization across all thesis chapters. Four interconnected strategies emerged: "AI as Assistive Tool, AI as Intelligent System, AI as Cognitive-Affective Partner, and Delegated Autonomy." Students reported leveraging AI from Chapter 1 through Chapter 5 for diverse purposes including "refining research ideas, generating outlines, assisting with literature searches, drafting methodologies, organising data, and structuring discussions." Notably, students maintained strong autonomy, with the majority allowing "only 50% or even less in the use of AI in their thesis completion," as Don stated: "I think it's about 50-50. Fifty is for.. to just spark the ideas and the other 50 and the process itself I am writing it."

Reports effect sizes.

Research paradigm

Interpretivist/Qualitative

Author conclusions

The authors conclude: "Findings reveal that students utilise AI as assistive aids and cognitive partners, with the central strategy of delegated autonomy emerging as key. This highlights a conscious choice to entrust specific tasks to AI while maintaining ultimate authorial responsibility." They assert that "students' commitment to critical evaluation and 'further checking' underscores a human-centred approach, proving that technological integration enhances intellectual scholarly work." The authors recommend that "universities must move beyond basic tool instruction to integrate explicit AI literacy into curricula, prioritising critical thinking and Delegated Autonomy."

Risk of bias

Selection bias: Purposive sampling of students with prior AI experience may not represent all undergraduate writers; Recall bias: Retrospective think-aloud protocol depends on accurate memory of events that occurred over several months; Social desirability bias: Participants may overstate their critical evaluation practices when discussing AI use; Interviewer bias: Open-ended prompts may be influenced by interviewer expectations; Selection bias: Purposive sampling of students who had voluntarily agreed to participate and had prior experience using AI tools; Recall bias: RTA protocol relies on retrospective recall of AI use over several months of thesis writing; Social desirability bias: Participants may have reported more responsible AI use in interviews than actual practice; Limited sample diversity: Only 8 participants from one institution in one region; Gender imbalance: 5 males and 3 females; Selection bias: Purposive sampling of participants who had already completed theses and agreed to participate, introducing self-selection bias; Recall bias: Retrospective reporting of AI usage decisions and strategies after task completion, potentially leading to inaccurate or reconstructed accounts; Small sample size (n=8) limits representativeness and statistical generalizability; Single institution studied (public university in central Malaysia), limiting geographic and institutional diversity; Language and translation bias: Some interviews conducted in Malay were translated to English, potentially losing nuance; Researcher bias in qualitative coding and theme interpretation, though iterative review is mentioned

Limitations

The authors acknowledge that "Although the sample size is small, it aligns with qualitative research principles that prioritise depth over breadth." The study is limited to "final-year undergraduate students" at "a public university in the central region of Malaysia" enrolled in "Arabic and English language studies," which constrains generalizability
The RTA method relies on retrospective self-report, which may not fully capture real-time cognitive processes
No concurrent think-aloud protocol was employed due to thesis writing being "an extended activity completed over several months, making concurrent verbalisations (as in standard TAP) impractical."

Open questions raised

Insufficient evidence on how students strategically delegate tasks to AI tools
Limited knowledge on how students filter or evaluate AI outputs
Lack of understanding of how AI use varies across thesis chapters
Gap in understanding how students maintain autonomy and academic integrity when working with generative systems
Limited empirical studies examining how students apply AI literacy principles in authentic thesis-writing contexts
The authors identified gaps in existing research on AI use in academic writing: "Much less is known about the specific processes by which students use AI during extended and high-stakes writing tasks such as in their undergraduate theses." They noted that "empirical studies examining how students apply these principles in authentic thesis-writing contexts remain limited" and that there is "insufficient evidence on how students strategically delegate tasks to AI tools, how they filter or evaluate AI outputs, how AI use varies across thesis chapters, and how students maintain autonomy and academic integrity when working with generative systems."

Extracted from: pdfAgreement 64%

Explore related topics

AI Manuscript Writing Academic Writing Qualitative