Using AI for data analysis: reflections on academic research practice

This blog contributes to a series of reflections on the use of Large Language Model (LLM) AI in academic research. Building on prior work examining AI in peer review (Bloomfield and Comfort, 2025) and conference preparation (Mooney and Bloomfield, 2026), it turns to the use of AI in data analysis. Here, I focus on how Microsoft Copilot (AI) can support the analysis of multiple interview transcripts for an internal report.

My involvement in the project

As part of a virtual internship opportunity within The Open University, I was given the opportunity to participate in a qualitative research project that explored the experience of delivering an apprenticeship programme within The Open University. The interviews were conducted by other members of the project team, and my role involved anonymising the interview transcripts, analysing them to identify key themes raised by participants, and selecting representative quotes to illustrate those themes. The findings contributed to an internal report, which provided an appropriate context in which to explore the use of AI in this way.

Using Microsoft Copilot to analyse interview transcripts

Managing multiple transcripts in Copilot

After anonymising the transcripts, I tried uploading as many as possible on Microsoft Copilot. However, I noticed that only three files could be uploaded at the same time (based on the software version made available to me by the University). Then, I uploaded all twelve transcripts in groups of three and asked, “Could you rank the interviewees from those who spend the most time on teaching tasks to those who spend the most time on operational tasks?”. Some interviewees have different roles in which some may not have teaching tasks; therefore, I wanted to know who had the most and the least teaching tasks.

I noticed that even after uploading the transcripts in groups of three, Copilot would “forget” the first six files I uploaded. Consequently, it only created the ranking for the last six and responded “no transcript uploaded in this chat” referring to the first six transcripts. Therefore, I performed the analysis by questioning Copilot for each group of six transcripts.

Using prompts to explore themes, roles, and emotions

Rather than relying on a single question, I learnt that analysing qualitative data with an LLM AI required experimentation. I used a series of prompts to explore how interviewees described their teaching, operational, and compliance responsibilities and how they felt about them. During this analysis, I realised that the way I phrased the questions would impact the responses Copilot would generate. Questions that were too broad often led to generalised responses, whereas questions with more details, consequently, led to more specific answers.

Additionally, using AI to analyse multiple files helped me identify potential themes, emotions and feelings across departments and roles, reducing the overwhelm of manually reviewing numerous transcripts.

Asking Copilot for direct quotes

When exploring these different themes, I also requested Copilot to provide direct quotes from the transcripts. I noticed that although the transcripts contained repetition and fillers such as “yeah”, “oh”, and “like”, Copilot removed these elements and ‘cleaned up’ the quotes it presented. I also observed that some quotes provided by Copilot combined separate parts of the same interview into a single continuous statement. In cases where an interviewee’s ideas were expressed across multiple turns, or developed in response to others, the AI would merge these fragments into one seamless quote instead of indicating pauses or interruptions (such as through ellipses). While adjustments created a more fluid narrative, it raised concerns about whether the structure and context of the original conversation were being accurately preserved.

Sensitivities and ethical considerations

Working with AI as part of a research project made me aware of the sensitivities surrounding its use. Because this was my first time engaging with qualitative data in this way, I questioned what counted as appropriate AI use for research. AI was undeniably helpful when handling multiple transcripts at once, but there were moments when Copilot would give answers to my questions without considering that the interviewees had different roles. This raised concerns about whether I was allowing it to shape the interpretation more than it should.

I also became more conscious of potential bias in AI-generated suggestions. Seeing patterns appear so confidently on the screen made it tempting to accept them, yet I knew that doing so would risk overlooking alternative interpretations. Understanding this helped me draw a clearer line between AI offering prompts and me making the final analytical decisions.

Conclusion

Reflecting on this experience, I have come to see AI not as a shortcut in the research process but as a tool that requires careful, thoughtful, and transparent use. As this work contributed to an internal report rather than an external publication, it provided a context in which the use of AI could be explored, while also raising ongoing questions about its implications for published research. 

Working with Copilot allowed me to engage with qualitative data in a way that felt both accessible and challenging, especially as a first-time user navigating unfamiliar academic territory. While AI helped me organise large volumes of information and spot early patterns across multiple interviews, the process also highlighted its limitations, particularly the risk of oversimplifying perspectives or overlooking important contextual differences between interviewees.

Jenifer Oshima

Jenifer graduated from The Open University in June 2025 with an Open (Hons) degree. 

From December 2025 to June 2026, she completed a virtual internship with the University, where she gained experience supporting qualitative research projects, coordinating academic events, and providing administrative support.