This one started as a conversation with Jessica: AI use is becoming ordinary in education, so... what if students disclosed it the way researchers cite the work they build on? Which opens the real questions — how do students and instructors actually read an AI disclosure, do they even want one, and what can it honestly tell you about a student's thinking?

Hours deep in thematic analysis — my own transcripts and my collaborators' — I kept hitting the same wall: spreadsheets do too little, professional analysis tools take too long to learn and only do what they were built to do. So I started wondering... what would AI help look like if it helped you organize, compare, and question your interpretations, without coding your data for you?

This one began mid-systematic-review, watching the AI literature grow faster than any team could hope to screen it. Somewhere in the batch-screening I caught myself asking: how much of this could a tool actually take on — faster, more accurate — without quietly giving up the traceability and rigor that make a review systematic in the first place?

AI hands you a polished solution in seconds — which is exactly the problem when you're trying to learn. The skill students actually need is metacognitive: knowing what the AI will do before you even ask. With collaborators at CMU and the University of Michigan, I'm building a scalable tutoring system that teaches students to write pedagogically sound prompts... asking for the help that helps them learn, not just `help me solve this problem`.

Months into using Codex and Claude Code on large projects, an agent once told me the file I'd asked it to change was "apparently not in use" (?!). Moments like that expose the gap between the agent's mental model of the codebase and mine — a gap that widens as projects grow with AI and we understand less and less of our own code. Can coding tools make that gap something you can actually inspect?

We gave a 700-student intro programming course an LLM assistant for a full semester and watched what happened — 8,000 uses, student surveys, educator interviews.