Most papers on AI in education stop at “should we use it” or “what students think about it.” Songkram, Chootongchai, Keereerat, and Songkram’s (2025) study in Interactive Learning Environments does something more interesting. It uses ChatGPT to do actual research work, generating code, summarizing literature, and refining questions, then documents what worked and what failed. Two studies happen in parallel within the same paper. The first analyzes survey data from 3,860 Thai primary and secondary students on what predicts their innovative thinking skills. The second is a transparent case study of using ChatGPT 3.5 across the full research lifecycle of that first study. When I read the paper in 2026, with two more years of AI development behind us and far more capable models on the market, the findings still hold up in instructive ways.
What the Student Survey Found
The substantive thread of the paper runs a regression model on three predictors of innovative thinking: digital technology skills, media and information literacy, and computational thinking. All three came out as significant. The strongest predictor was computational thinking, with a regression coefficient of 0.38, followed by media and information literacy (0.22) and digital technology (0.21).
The result worth noting is that computational thinking carries the most weight in predicting innovative thinking skills among Thai students. For every one-point increase in computational thinking on the survey scale, innovative thinking went up by 0.38 points. The authors take this as a clear signal that schools should prioritize computational thinking instruction, including coding, algorithmic reasoning, and problem decomposition.
ChatGPT as Research Assistant
The methodological thread is what I find most useful in this paper. The authors used ChatGPT 3.5 at multiple stages: idea generation, literature summarizing, methodology selection, R code generation for regression, and interpretation of results. They reproduce actual transcripts of their dialogues with the model, which makes the workflow transparent in a way most AI-assisted papers aren’t.
Some uses worked well. ChatGPT generated working R code for the regression analysis. It suggested appropriate statistical methods given the dataset. It helped summarize literature and synthesize related work into a manageable shape for the writing-up phase.
Other uses failed in ways worth taking seriously. The fabricated citations problem the authors document directly is the one I’d flag first. This connects to Butson and Spronken-Smith’s (2024) earlier work on AI implications for research in higher education, where they raised similar concerns about source verification.
The Limitations Are the Real Story
Three limitations the authors document are worth highlighting.
First, hallucinated citations. The authors note that “ChatGPT autonomously produces text without referencing actual publications, [and] any citations it generates upon request are entirely fictitious” (p. 1706). They documented this directly in their own use of the tool. Every citation the model produces has to be verified manually, which eats the time savings AI was supposed to provide.
Second, generic content. The model produces output that sounds reasonable but lacks topic specificity. For research on niche or culturally embedded subjects, the authors found ChatGPT’s responses too general and frequently inaccurate. The output reads as if it was assembled from broad patterns in training data, which is exactly what it was.
Third, contextual blind spots. The authors describe how the model “struggled with nuanced, context-specific interpretations of our data, particularly regarding cultural factors influencing innovative thinking in Thai education” (p. 1707). This is the limitation I think researchers underweight the most. AI tools trained on largely English-language, Western-dominant corpora don’t have the cultural fluency needed for cross-cultural research interpretation.
I’ve covered related concerns in my post on Cheng, Calhoun, and Reedy’s (2025) recommendations for AI-assisted academic writing, where the authors arrive at similar conclusions. AI is useful for some research tasks, dangerous for others, and the line between them is the writer’s own judgment.
What This Means Going Forward
Songkram et al. (2025) recommend a hybrid model. ChatGPT can speed up code generation, brainstorming, and literature summarizing. Humans retain control over interpretation, citation verification, ethical decisions, and contextual reading. The authors argue for transparent reporting of AI use at every stage of research, which lines up with what Cleland et al. (2025) called for in their work on AI disclosure in academic publishing.
The most useful contribution of this paper, for my reading, isn’t the regression finding about computational thinking, important as that is. It’s the documented record of what ChatGPT could and couldn’t do in a real research project. The technology has moved forward since this study was conducted, but the categories of failure the authors identify, including hallucination, genericness, and missing cultural context, are categories of limitation, not just technical glitches with the 2024 version of the model. They’ll persist as the technology develops, even as the surface forms change.
For researchers thinking about how to integrate AI into their work, this paper is one of the better starting points I’ve seen. Read it for the empirical findings about Thai students. Read it again for what it teaches about working with AI as a research partner.
References
Songkram, N., Chootongchai, S., Keereerat, C., & Songkram, N. (2025). Potential of ChatGPT in academic research: Exploring innovative thinking skills. Interactive Learning Environments, 33(2), 1689-1711. https://doi.org/10.1080/10494820.2024.2375342
Rank Math SEO Data:
- SEO Title: ChatGPT as a Research Assistant: A Case Study from Thailand
- Meta Description: What a Thai study using ChatGPT 3.5 across the research lifecycle documents about hallucinated citations, generic content, and missing cultural context.
- Focus Keyword: ChatGPT in academic research
- Secondary Keywords: ChatGPT research assistant, AI in research, computational thinking innovative thinking, ChatGPT hallucinated citations, ChatGPT limitations research, Songkram ChatGPT, AI cultural context research
- URL Slug: chatgpt-research-assistant-case-study
Suggested Category: AI Policy & Ethics
Word count: ~950 words