July 6, 2026

Abstract

Large language models (LLMs) are increasingly embedded in students’ academic work. They can support tasks such as brainstorming, summarizing literature, drafting essays, and technical work. Yet students may misjudge when, how often, and for which purposes they rely on these tools.

In this work, the authors introduce PromptMirror: a student-facing dashboard that helps learners reflect on their historic LLM conversations. PromptMirror visualizes personal interaction patterns to encourage reflection on LLM reliance and learning practices.

The Study

The PromptMirror research was conducted by Ka Hei Carrie Lau, Dr. Nađa Terzimehić, and Prof. Dr. Enkelejda Kasneci at the Chair of human centered technologies for learning and the Munich Center for Machine Learning.

The project included two complementary focus groups: one with four experts from education, cognitive science, HCI, and privacy research, and one with four student users of LLMs. Together, they identified metrics and design features that could support reflection on LLM use. Students highlighted the importance of information about usage patterns, topics and tasks, and the quality and effectiveness of LLM interactions.

These insights informed PromptMirror’s design space. Students can upload their historic LLM conversation data and explore four reflection-oriented perspectives:

• Usage statistics, showing temporal and frequency patterns;
• Sentiment analysis, comparing the sentiment of user prompts and AI responses;
• Intent analysis, identifying purposes such as technical, educational, or creative use; and
• Thematic analysis, using word clouds to highlight recurring topics.

To examine how the dashboard may support reflection, the researchers conducted online think-aloud sessions with 20 university students. Participants uploaded and explored their own historic LLM interaction data while sharing their observations and reactions.

Key Findings 

Making an estimation gap visible
Some participants noticed a discrepancy between how much they believed they relied on LLMs and what their data showed. The authors describe this as an estimation gap: a tendency to misjudge one’s own LLM use. The dashboard made these patterns more visible and prompted participants to reconsider their personal reliance on AI tools.

Moving beyond simple usage counts 
Participants did not only comment on how often they used LLMs. Their reflections connected the visualizations to broader learning practices, including task-specific uses of AI, trust in AI-generated answers, and concerns about reliability.

Supporting deeper reflection on AI use
Participants reflected on their LLM-use practices, including prompt engineering, alternative resources, privacy-preserving practices, and questions of responsibility and accountability. Some also articulated goals for changing or improving aspects of their future LLM use.

A useful starting point, while behaviour change remains open
Participants generally found PromptMirror easy and enjoyable to use for reviewing their long-term LLM-use history. However, the study examined a one-time interaction and did not measure sustained behavioural change. Further research should investigate whether reflective analytics can support continuous reflection and more self-regulated LLM use over time.

Takeaway

PromptMirror offers students a way to look back on their own LLM-use history and make otherwise invisible usage patterns easier to see. By visualizing when, how often, and for what purposes students use LLMs, the dashboard can support reflection on reliance on AI in learning, trust in AI outputs, prompting habits, privacy, and alternative learning strategies. The work provides preliminary evidence that reflective visualizations can help students become more aware of their everyday LLM use. Future versions could extend PromptMirror with reflective prompts, richer metrics, and longer-term support for continuous reflection, both after and during LLM interactions.

Try Promptmirror: https://promptmirror.edu.sot.tum.de/

References

• Ka Hei Carrie Lau, Nađa Terzimehić, and Enkelejda Kasneci. 2026. PromptMirror: Visualizing LLM Use to Support STEM Student Reflection. In Proceedings of the 2026 Designing Interactive Systems Conference (DIS ’26). Association for Computing Machinery, New York, NY, USA, 1886–1906. https://doi.org/10.1145/3800645.3812863