Architect Mode
Why I keep AI out of my students’ hands — and use it to build personalized teacher tools instead.
I consider myself fairly tech-savvy. I always have been, and my students usually are too. But for most of the last year, I’ve been quietly resistant to putting AI in front of them. Lately, the more I watch eighth graders work, the more I see a pattern that predates ChatGPT but has been supercharged by it: the pull toward finishing faster, toward smoother text, toward the right answer delivered rather than built.
So when AI arrived in classrooms promising personalized support and instant feedback, I found myself quite hesitant about integration. This post is my attempt to name that hesitation and to describe where I’ve landed instead.
What the research is starting to say
The evidence on student AI use is still early, but it is pointing in a direction worth taking seriously. A 2025 systematic review on AI and critical thinking found a significant negative correlation between frequent student AI use and analytical reasoning, with cognitive offloading (letting a tool do thinking the student would otherwise do) as the mediating mechanism (Gerlich 2025). The same review pointed to metacognitive monitoring as one of the casualties: the more students lean on AI, the less they practice the internal habit of checking their own thinking, and the effect compounds for learners who haven’t yet built foundational skills in the discipline. Younger learners in the review showed higher AI dependence and lower critical thinking scores than older learners. Developmentally, the students most at risk are the middle schoolers I teach.
A 2025 MIT preprint by Kosmyna and colleagues adds another piece. University students who used ChatGPT to write essays showed measurable reductions in neural connectivity associated with deep cognitive engagement, compared to peers writing without AI assistance. The sample was small, and the design has limits I don’t want to overstate, but the direction of the findings tracks with what teachers have been observing on the ground.
There is one counterintuitive finding worth holding onto. A three-wave study by Wang and Zhang, with over nine hundred students, suggests that how students orient toward AI matters more than whether they use it. Students who treated AI as a thinking partner, questioning its outputs, building on them, and comparing them to their own ideas, showed stronger transfer learning than students who used it as an answer machine. The trouble is that a thinking-partner orientation requires metacognitive habits most thirteen-year-olds are still building, and the default orientation tends to be the answer-machine one.
When I sit with all of this, I land somewhere honest. The version of AI use that benefits students is sophisticated, and the version available to them by default is not, which is why I’m not in a hurry to put it in their hands.
So, where does AI fit in my classroom?
It fits behind me, rather than in front of me.
I’ve started calling this architect mode. The teacher uses AI in the design phase, building the materials, the scaffolds, and the protocols, and keeps it out of the student-facing layer. The AI doesn’t deliver instruction, tutor, or grade. It helps me build the thing my students will then think their way through, by hand, at the speed of their own minds.
Architect mode draws a different line than the usual “AI yes, AI no” debate I keep seeing in teacher spaces. The line I’m drawing isn’t whether AI is in the classroom; it’s where AI sits in the workflow. Behind the teacher, it works as a design tool. In front of the student, it tends to become a cognitive shortcut. The same technology can pull in two very different directions depending on where you place it.
What architect mode produces: digital scaffolds
The artifacts that come out of architect mode aren’t ed-tech products. They’re something smaller and more specific, and I’ve started calling them digital scaffolds.
The distinction matters more than it sounds. Ed-tech products are pre-built, sold to schools, and tend to assume your context for you. They’re polished, expensive, and often rigid in ways that ask you to bend your teaching around them. Digital scaffolds are scrappy, free, and purpose-built for a specific unit, a specific class, and a specific pedagogical goal. The tool bends around the teaching, rather than the other way around.
No ed-tech vendor was going to build the tool I needed for my dystopian book club unit, so I built it myself.
A case study: the Power EKG
We’ve been reading dystopian novels in book clubs, including The Hunger Games, Divergent, Maze Runner, Feed, and The Giver, and tracking how the protagonists’ power rises and falls across the narrative. I wanted students moving past plot summary into structural analysis, sitting with questions about what kind of power is shifting, what causes it, and the question I most wanted them holding: Does the system allow this shift, or does it absorb it?
I built a tool I called the Power EKG Builder. Students plot three to five power shifts on a chart, and for each shift, the tool requires them to choose a power category from four options (political, information, social, narrative), identify the cause, articulate what was gained or lost in terms of the specific power rather than the event itself, and answer whether the system allowed the shift or absorbed it.
Every design choice in that tool is what I’ve been calling friction-by-design. The four power categories force a typology before students can write anything. The required fields keep them from skipping the analytical move and writing a plot summary instead. The system-absorption question is genuinely hard, and the tool tells students that explicitly, so they don’t feel stupid for sitting with it.
The part I want to be honest about is the division of labor. AI built the interface, while I designed the thinking architecture. The categories, the questions, the order they appear in, and the wording that distinguishes “what happened” from “what power was lost” are all pedagogy, and they came from me. AI gave me the speed to ship a working tool in an evening that would have taken me weeks to build by hand, if I could have built it at all.
A note about my dysgraphia students
For most of my students, the Power EKG adds productive friction. It slows them down and forces precision, which is what I want.
For a student with dysgraphia, the calculation changes. The mechanics of handwriting are already a heavy cognitive load that competes with, and often drowns out, the analytical thinking I’m trying to surface. On a paper graphic organizer, that student is spending cognitive budget on letter formation, legibility, spacing, and pencil pressure long before they reach the analysis itself. The friction they’re experiencing isn’t the kind I’m trying to create; it’s the kind I need to clear out of the way.
Some of my dysgraphia students don’t enjoy being singled out with accommodations; they want to use the same tools as the rest of their peers. A well-designed digital scaffold removes the unproductive friction (typing, structured fields, saved progress) so the productive friction can finally happen. My dysgraphia students can think about power instead of about handwriting, and the same tool ends up doing opposite work for different students while serving the same pedagogical goal.
This is the quieter case for digital scaffolds, and the one I didn’t expect to find myself making. Inclusive design and friction-by-design aren’t actually in tension. A good scaffold calibrates friction, adding it where thinking needs to slow down and removing it where it’s eating the cognitive budget meant for something else.
Where I’m landing
I’m still not putting AI in students’ hands. The research gives me real reasons to wait, and my own values give me more.
If I’m honest, this whole project is continuous with the work I’ve been doing on this Substack from the start. The metacognitive habits I’ve been trying to build all year, the cognitive wobble, the second-guessing, the willingness to sit in not-knowing, are precisely the ones the research suggests AI use erodes fastest. Architect mode is, in a way, my attempt to keep building those habits in students while still letting AI do the work it’s actually well-suited for, which is helping me design better learning experiences.
What architect mode has changed is what I can make. I’m building digital scaffolds for other units now, including a claim-evidence-reasoning coach, a structured protocol for lateral reading, and a discussion organizer that pushes students to cite before they react. None of these would have been feasible a year ago, and all of them exist because AI lets me design according to the needs of my classroom.
Maybe the most useful move for a teacher right now isn’t adopting the latest AI platform. It might be slipping into architect mode for an hour on a Sunday afternoon and building the small, specific, deliberately frictional thinking tool that this week’s lesson actually needs.
* * *
A note on what’s next: I’m working on a manual for teachers who want to learn this mode, including how to design digital scaffolds, the prompt patterns I use, how to host them with ease, and the working tools from my own classroom. I’m looking forward to sharing this soon.
* * *
Sources:
Kosmyna, N., et al. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt When Using an AI Assistant for Essay Writing Task. arXiv:2506.08872.
Wang, S., Zhang, H. Pedagogical partnerships with generative AI in higher education: how dual cognitive pathways paradoxically enable transformative learning. Int J Educ Technol High Educ 23, 11 (2026). https://doi.org/10.1186/s41239-026-00585-x
Gerlich, M. (2025). AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies, 15(1), 6. https://doi.org/10.3390/soc15010006