Most research on AI in education looks outward. Surveys of students, interviews with faculty, experiments comparing AI-assisted and non-assisted groups. Panke (2025) did something different. She spent five months (May to October 2024) documenting her own use of generative AI across research, teaching, and instructional design, then analyzed the entire record through the lens of activity theory. The result is an autoethnography that reads less like a typical empirical study and more like a field diary from someone trying to figure out, in real time, what this technology actually does to the way we work and think.
I find this kind of study genuinely useful because it doesn’t claim to know the answers. Panke produced 92 field notes and 93 AI queries during the observation period, coded them systematically, and mapped them onto the components of an activity system: subject, tools, rules, community, division of labor, and object.
Generative AI in Education Research: Tool, Subject, or Both?
One of Panke’s central observations is that AI kept shifting roles during her research process. Sometimes it functioned as a tool, mediating her work the way a search engine or a reference manager would. Other times it became the subject of her inquiry, the thing she was studying. And occasionally it acted almost like a collaborator, generating outputs that redirected her thinking in ways she hadn’t planned. Panke argues that “it is insufficient to define AI simply as mediating artifacts because of the generative ability to produce new activities and outputs and to structure activity beyond the subjects’ intentions” (p. 237).
That’s a meaningful claim. Most frameworks for understanding technology in education treat tools as passive mediators. You use a tool; the tool doesn’t use you. But generative AI blurs that line. I covered Butson and Spronken-Smith’s (2024) argument about how AI is reshaping the research process in higher education, and Panke’s autoethnography adds granular, first-person evidence to that argument. She documents moments where AI pushed her toward lines of inquiry she hadn’t considered and moments where it confidently produced garbage she had to catch because of her subject matter expertise.
Panke found it genuinely difficult to predict which tasks AI would handle well and which would be a waste of time. The largest share of her AI queries (41 out of 93) went toward data collection and analysis, the analytical work that should, in theory, benefit most from AI’s pattern-recognition abilities. Drafting, editing, and review accounted for smaller portions. Literature review, surprisingly, got only a single query.
What’s telling is the distribution across quality. AI was, as Panke puts it, “likewise impressive and error-prone in both higher-order thinking and automation tasks.” It wasn’t reliably good at the simple stuff and reliably bad at the complex stuff. The failures were scattered unpredictably, which meant Panke could never let her guard down.
Every output required verification, regardless of the task’s difficulty level. That tracks with research I’ve covered on AI and qualitative analysis by Anis and French (2023), who found that AI tools could accelerate certain analytical tasks but always required human oversight to maintain interpretive depth.
The field notes also reveal something that I think resonates with any educator who’s been using these tools seriously: the feeling of never having used AI “enough.” Panke describes a persistent fear of missing out when she didn’t use AI and a sense of inadequacy when the tool didn’t produce the results she expected. She writes that she “never gained a sense of ‘enough’ AI-input” (p. 241). That’s a psychological dimension of AI adoption that quantitative studies rarely pick up.
What AI Changed in Instructional Design
The instructional design findings are the most practically rich section of the paper. Panke catalogs a wide range of AI uses: generating course materials (objectives, outlines, scripts, discussion prompts), producing multimedia content (video libraries, case study simulations, audio with Suno), building quizzes and interactive assessments, creating accessible versions of content (transcriptions, audio PDFs, text descriptions of visualizations), and drafting documentation like faculty handbooks and training manuals.
The pattern across all of these uses is that AI accelerated production but did not reduce the need for professional judgment. Course materials generated by AI needed significant editing. Quiz items had to be checked for accuracy, and multimedia content came back with tone and audience problems that required manual revision. Panke’s field notes document a recurring tension: AI made it faster to produce a first draft of almost anything, but the time saved on drafting was often consumed by the editing and quality control that followed.
I wrote about Mishra, Warr, and Islam’s (2023) TPACK framework and how it positions pedagogical knowledge as the strongest predictor of effective AI integration. Panke’s experience confirms that claim from the ground level. Knowing how to use AI is secondary to knowing what good instruction looks like.
The most revealing finding comes at the very end of the paper, almost as an aside. Panke writes that despite frequent AI use across all her professional work, she “seldom or never” uses AI in her private life. The reason: “most of my output is not judged by its efficacy and efficiency, but by the joy in creating it” (p. 242). She adds that “the tools we use shape the way we think, the reverse is likewise true. The way we think, and our objectives for usage, shape the tool” (p. 242).
If AI is most useful when the goal is efficiency and least useful when the goal is creative satisfaction, then education is caught in an uncomfortable intersection. Learning is supposed to be about the process, about the struggle and the thinking that happens along the way. But the entire pitch for AI in education is efficiency: faster feedback, quicker content production, automated grading.
Shaw and Nave (2026) warned about cognitive surrender, the risk that students stop thinking critically because AI handles the reasoning. Panke’s personal finding suggests the risk extends beyond students. When the tool is optimized for output, the human gravitates toward output too. The creative process, the messy, joyful, inefficient part, gets sidelined.
Panke’s paper doesn’t prescribe solutions, and I respect that. She calls it “a spin down the AI rabbit hole” that documents what is, not what should be. For educators trying to figure out their own relationship with AI, that kind of documentation is what we need right now. The prescriptions can come later. First, we need to understand what’s actually happening in our own practice.
References
- Anis, S., & French, J. A. (2023). Efficient, explicatory, and equitable: Why qualitative researchers should embrace AI, but cautiously. Business & Society, 62(6), 1139–1144. https://doi.org/10.1177/00076503231163286
- Butson, R., & Spronken-Smith, R. (2024). AI and its implications for research in higher education: A critical dialogue. Higher Education Research & Development, 43(3), 563-577. https://doi.org/10.1080/07294360.2023.2280200
- Mishra, P., Warr, M., & Islam, R. (2023). TPACK in the age of ChatGPT and generative AI. Journal of Digital Learning in Teacher Education, 39(4), 235–251. https://doi.org/10.1080/21532974.2023.2247480
- Panke, S. (2025). How can (A)I research this? An autoethnographic exploration of generative AI in research, teaching and instructional design. Journal of Teacher Education, 76(3), 230-244. https://doi.org/10.1177/00224871251325065
- Shaw, S. D., & Nave, G. (2026). Thinking fast, slow, and artificial: How AI is reshaping human reasoning and the rise of cognitive surrender. Working paper, The Wharton School, University of Pennsylvania. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6097646