AI is now part of many university students’ study routines. But what do students actually do with it—and what can their recorded interactions tell us about learning?
Generative AI has prompted debate about academic integrity, assessment and independent thinking. A survey of students at four Australian universities provides broader context for that debate (Chung et al., 2026). Here, we take a closer look at a different, smaller dataset: anonymised records from a specific AI-supported university learning task, including student–chatbot dialogue, prior experience, and assessment outcomes (Li et al., 2026).
Scope matters: This is an observational study of participants in a particular task. It is not a representative sample of all Australian university students, and any relationships shown below are associations—not proof that using AI improves or harms grades.
In the recorded conversations, 43.6% of classified interactions involved evaluating ideas, compared with 22.4% about writing advice. These are shares of messages, not percentages of students. The prominence of evaluation-related questions suggests a broader role for AI in this task than drafting text alone.
The median participant with recorded dialogue sent 11 messages; the upper quartile sent 16 or more. A single average would conceal these differences. The dataset cannot establish how much time students spent thinking outside the chatbot.
This comparison adds a third dimension: question family, prior knowledge and share of interactions. It is descriptive; differences may also reflect the task, student background and varying levels of participation. Because prolific users contribute more messages, these proportions should not be interpreted as the proportion of students using each question type.
Each dot represents a participant with both dialogue and a recorded proposal score. The plot considers interaction volume, assessment score and prior AI knowledge together. It cannot tell us whether additional questions caused any change in performance: students’ starting abilities, task strategies and other factors may differ.
This chart compares perceived usefulness, self-reported improvement and the number of respondents. It does not measure actual gains in understanding. The post-task survey was completed by fewer students than the full study cohort, so nonresponse may affect the pattern.
In this task, AI use involved evaluating ideas, seeking writing advice and exploring possible approaches. Participation varied widely. The final three charts show how engagement can be examined alongside prior experience, grades and student perceptions, but none establishes that AI itself caused better learning or higher marks.
For universities, the practical question is not simply whether students use AI. It is how students use it, which forms of support they find valuable, and how institutions can assess learning fairly when access and study practices vary.
The five figures are original R visualisations built from the released dataset accompanying Li et al. (2026). Dialogue categories are researcher-assigned codes. Interactions are counted at the message level; student-level analyses use anonymised IDs. When a survey contains repeated IDs, this report retains the most recent timestamped response per student; where timestamps cannot be parsed, the first encountered response is retained. Missing survey answers are not replaced with zero. Analyses use available cases, so sample sizes vary across figures. No private message text is reproduced.
The broader Chung et al. (2026) study provides context but is not merged with the Monash task-level records. The findings here cannot be generalised to every Australian university or interpreted causally.
The original data were collected and released by me. OpenAI ChatGPT was used to help plan the narrative, and troubleshoot R code, and refine written explanations. The author is responsible for verifying the output against the dataset, checking the interpretations, reviewing the citations and submitting the final work. I used this to polish my grammar and correct it.
Chung, et al. (2026). The use and usefulness of GenAI in higher education: Student experience and perspectives. Computers and Education Open. https://www.sciencedirect.com/science/article/pii/S2666557326000182
Li, et al. (2026). Research dataset on generative-AI-assisted information problem solving in education. Scientific Data. https://doi.org/10.1038/s41597-026-08054-w
Research dataset accompanying Li et al. (2026). Figshare. https://doi.org/10.6084/m9.figshare.32321499
OpenAI. (2026). ChatGPT [Large language model]. https://chatgpt.com/