I’ve been using AI tools in my teaching for a full academic year now. Not detection tools, but generative and assistive tools on my end. Planning, differentiation, feedback drafts. I think it’s worth writing up what actually held and what didn’t.
What held: using AI to generate first drafts of rubrics I then revise. Generating differentiated reading versions of complex texts. Drafting initial feedback on student work that I then substantially edit and personalize. All of these saved time in ways I could actually measure.
What didn’t: using AI to generate discussion questions. The questions were always slightly off, too broad or missing the specific thing I wanted to surface from the text. I gave up on this after a few months. Also: using AI to analyze student writing patterns across a class. The summaries were generic and missed what I actually noticed.
The pattern is that the tools work when my judgment is the final step and the tool is handling something earlier in the process. They don’t work when the tool is supposed to replace the judgment.
Curious if this tracks with what others teaching with these tools have found.
The measurable time savings point is important and underrated. Most AI tool evaluations are vibes-based. ‘I feel like it saves time’ isn’t the same as actually tracking it. The fact that you could measure it is a sign you’re using the tool for the right things. Vague productivity gains are usually a sign the tool is filling time rather than saving it.
The discussion question finding resonates. Good discussion questions are highly specific to what happened in that particular class, with that particular text, with those particular students. That specificity is exactly what AI can’t access. It doesn’t know what you need to surface. You do.
The ‘judgment as final step’ principle is a good general heuristic for AI tool adoption. It applies outside education too. The tools that integrate cleanly are the ones where the human is still clearly in the decision seat. The ones that try to make the decision produce either low adoption or bad outcomes.
from the student side, the differentiated reading thing genuinely helps. like when the easier version exists i actually understand the hard version better afterward. it doesn’t feel like dumbing down, it feels like having a starting point. so from where i sit that one is working
This tracks very closely with my experience. The tools that require me to make the final call are the ones I kept using. The tools that were supposed to generate the finished product dropped out of my workflow within a term. The distinction between ‘generates a draft I shape’ and ‘generates a deliverable’ is where the value actually lives.