I first discussed AI disclosure with Jessica through an analogy. If AI had become ordinary in education, could students acknowledge it the way researchers cited work they built on? That conversation led us to examine how students and instructors interpreted these disclosures, whether either group wanted them, and what a disclosure could honestly reveal about a student's thinking.

My own thematic analysis, working alone and with collaborators, kept exposing the same mismatch. Spreadsheets did too little, while professional tools took too long to learn and forced the analysis into workflows that never felt like mine. While designing an alternative, I treated AI as a partner around the interpretation rather than the one doing the interpretation. For me, the personal act of making meaning lay at the heart of qualitative research. The system organized, compared, and surfaced material while leaving that act with the researcher.

I experienced the scale problem firsthand during a systematic review. The AI literature grew so quickly that the corpus changed even while we screened it. As I batch-processed papers, I asked where a tool could save time and improve accuracy without weakening the traceability and scientific rigor that made the review systematic.

I worked with collaborators at CMU and the University of Michigan on a scalable tutoring system built around a paradox in computing education (really in any AI-assisted learning+. AI produced polished corect solution in seconds, but a perfect solution is not going to be helpful to someone still trying figure out how everything works. We treated prompting as an AI-literacy and metacognitive skill. Students practiced anticipating what a model would do before asking, then wrote prompts that elicited guidance rather than finished answers.

My extended use of Codex and Claude Code on large projects repeatedly exposed discrepancies between the agents' view of a codebase and my own. At one point, an agent confidently proposed changing a file that was not even in use. Those moments made the mental-model gap concrete and motivated CopilotLens, which exposed the context behind generated changes so developers could inspect them instead of letting their understanding of the codebase erode.

I helped deploy an LLM assistant in a 700-student introductory programming course for a full semester. Across 8,000 uses, student surveys, and educator interviews, we observed where AI support encouraged novices to keep reasoning and where it let them bypass that work.