- The paper reveals that LLMs are used beyond mere text correction, aiding in conceptual development and methodological structuring.
- Students disclosed their LLM usage to foster transparency, critically balancing improvements in clarity with risks like hallucinations and semantic distortions.
- Empirical findings suggest that structured disclosure of LLM assistance can promote responsible, reflective, and rigorous academic writing practices.
Analysis of LLM Usage in Software Engineering Student Research Writing
Integration of LLMs in Undergraduate Empirical Assignments
The reported educational experience centers on the explicit integration of LLMs within a third-year undergraduate software architecture course assignment, where students applied either rapid review or grey literature review methodologies to synthesize evidence and communicate research findings. Notably, the pedagogical design required students to disclose their LLM usage, fostering transparency and critical reflection regarding AI-assisted academic work. The disclosed reflections from 146 students were systematically analyzed, utilizing both LLM-assisted excerpt categorization and manual verification, capturing authentic patterns of technology appropriation.
Categorization of LLM Utilization
The analysis revealed a multi-faceted deployment of LLMs across the writing workflow, distinguished by five key dimensions:
- Presentation: The dominant mode of LLM usage pertained to text refinement—grammar correction, sentence restructuring, formatting, and adaptation to academic conventions. Students frequently described LLMs as "final editors," routinely leveraging tools such as ChatGPT, Gemini, and Grammarly.
- Novelty: Students appropriated LLMs for initiating academic reasoning: brainstorming research topics, narrowing investigative scope, and elucidating unfamiliar concepts, indicative of LLMs functioning as dynamic idea generators and conceptual clarifiers.
- Relevance: Several disclosures highlighted LLMs as facilitators in organizing, articulating, and synthesizing findings, effectively improving the communicative clarity and coherence of empirical arguments.
- Rigor: LLMs contributed to methodological structuring, including clarification of review protocols, formulation of research plans, evidence categorization, and empirical procedure description, underscoring their potential as supportive guides for research organization and process conceptualization.
- Transparency: Students reported using LLMs to enhance explanation, traceability, and reproducibility of research procedures, albeit less frequently. The reflective disclosures suggest emerging student awareness regarding methodological clarity in academic communication.
These usage patterns underscore that student engagement with LLMs extends beyond superficial language corrections, encompassing cognitive and structural support for empirical academic tasks.
Reported Benefits and Challenges
Students articulated distinct benefits derived from LLM usage: improved grammar and style, facilitated organization, enhanced comprehension of complex concepts, and reduced writing difficulty. LLMs were perceived as interactive partners in both the generation and expression of academic content. However, students consistently identified major risks, including hallucinations, fabricated outputs, semantic distortion, and unreliability. Explicit examples stress concern over content accuracy and preservation of intended meaning.
There is evidence that students engaged critically with LLM-generated suggestions, verifying, revising, and filtering outputs to retain authorship and integrity. The reported experience did not reflect uncritical acceptance; rather, it signaled evolving student strategies to balance utility and reliability—echoing broader pedagogical debates around AI inclusion and responsibility in academic workflows.
Educational Implications and Lessons Learned
The empirical analysis of student disclosures generates several actionable insights for integrating LLMs in software engineering education:
- Introducing disclosure requirements encourages transparency and critical engagement, compelling students to self-assess both benefits and risks.
- LLMs serve as companion tools, not mere correction mechanisms, supporting both conceptual development and methodological organization within empirical assignments.
- Risks of AI-generated content (e.g., hallucination, misalignment, semantic distortion) persist, demanding explicit verification protocols and critical reflection exercises.
- The educational context shapes student strategies; assignment design and institutional guidelines may determine the depth and quality of reflective LLM usage.
- There is opportunity to further research how disclosure-driven reflection affects learning outcomes, authorship confidence, and academic integrity across broader contexts.
Limitations
The study's analytic scope is limited to a single undergraduate course and self-reported disclosures after assignment completion. Generalizability may be constrained; the depth and fidelity of reflections are affected by the voluntary, post-hoc nature of the data. The absence of direct interaction or usage logs limits insights into real-time LLM engagement modalities. Future research utilizing interviews, observational data, and comparative course designs would enhance external validity and granularity of findings.
Implications for Future AI-Integrated Pedagogy
The findings illustrate that students are actively experimenting with multiple LLMs and developing workflows that combine human judgment with AI assistance. Practical implications include a need for assignment protocols that mandate reflective disclosure, guidelines addressing the risks of hallucination, and explicit strategies for evidence verification. Theoretically, the study calls for further exploration of AI-mediated learning processes, responsible authorship negotiation, and meta-cognitive skill development in empirical software engineering education.
Anticipating future developments, expansion of empirical studies to encompass systematic and multivocal literature reviews, as well as more diversified course contexts, will yield richer understanding of AI's transformative role in academic evidence synthesis, methodological reasoning, and communication in software engineering pedagogy.
Conclusion
This experience report provides an authoritative characterization of how software engineering students appropriate LLMs in evidence-based academic writing. The disclosed patterns indicate systematic deployment of LLMs beyond linguistic support, contributing to idea generation, methodological clarification, and transparency in empirical assignments. Students’ reflective disclosures elucidate ongoing negotiation between utility and reliability, underscored by concerns for authorship and content verification. Integrating structured disclosure and critical engagement will be instrumental in evolving AI-assisted educational practices, laying a foundation for responsible, transparent, and rigorous adoption of LLM technologies in empirical software engineering education.