Researchers cite the work they build on. It's just part of how we write. So when AI became an everyday part of learning, I kept wondering: could students acknowledge it the same way? I brought the idea to Jessica, and we started digging into it together. But a disclosure is a strange thing to ask for. What does “I used ChatGPT” actually tell an instructor about how a student thought? Do students want to write one, do instructors want to read one, and if being honest has a cost, who ends up paying it?

Qualitative analysis means sitting with your data: rereading transcripts, moving quotes around, arguing with collaborators about what someone really meant. The tools never fit that for me. Spreadsheets do too little, and professional tools take forever to learn and push you into workflows that don't feel like your own. AI can now code a whole dataset in minutes, but then whose analysis is it? What if AI helped around the interpretation, organizing material and surfacing what you missed, and left the meaning-making to you? ScholarMate tries this for one person making sense of their material, and Reflexis brings it to teams working through disagreements.

A systematic review is meant to be the careful, reproducible way to read a field. Then I tried doing one on AI, and the field kept growing while we were still screening. Most of the work is toil: abstract after abstract, the same inclusion criteria applied over and over. That sounds like a perfect job for an LLM. But the whole point of a systematic review is that every decision can be traced and checked later. So which parts can a tool take over, and which parts should stay with the researcher?

Ask an AI for help on a CS1 assignment and you'll get a polished, correct solution in seconds. That's handy if you already know how to code. If you're still figuring out how a loop works (or learning anything, really), a perfect answer doesn't teach you much. With collaborators at CMU and the University of Michigan, we started treating prompting as a skill students can learn, part AI literacy and part metacognition. Students practice predicting what the model will do before they ask, then write prompts that get them guidance instead of finished code. We later tested this in an RCT in a real CS1 course.

Coding agents can now change a lot of code in one go, but they rarely tell you what they were thinking. I use Codex and Claude Code on large projects, and every so often I notice the agent's picture of the codebase is quite different from mine. Once, an agent set out to edit a file that wasn't even in use. If I stop checking, how long before my own picture of the code starts to drift too? CopilotLens tries to make that gap visible. It shows the plan behind an agent's changes and the parts of the codebase that shaped them, so you can check its reasoning instead of just accepting the diff.

What happens when 700 students in an intro programming course get an AI assistant for a whole semester? I helped deploy CodeAid, which was designed to help without handing out solutions, and students used it about 8,000 times. Between those logs, student surveys, and interviews with educators, we could see where the AI kept novices reasoning and where it let them skip the reasoning entirely.