AI that helps kids think, not think for them.
A two-month, end-to-end design research project reimagining how AI shows up in K-12 classrooms: adapting to how each student learns and feels, without eroding the critical thinking they're still building, and earning the trust of the teachers who have to stand behind it.
What it looks like.
Example of screens from each side of the product
The problem isn't that students use AI. It's that most tools jump straight from prompt to answer.
Offloading a low-level task to AI is fine, we do it with calculators and spellcheck. The danger is cognitive surrender: "help me improve my argument" becomes "write it for me," until students stop asking is this right? and start asking what did the AI say? Good AI UX should route through thinking.
R1 · The student side
How can AI adapt to each student's learning style and emotional needs in a way that supports their thinking, without replacing the critical thinking and core skills they still need to develop?
R2 · The teacher side
How can AI genuinely aid teachers in the classroom while addressing their trust and comfort with the technology, rather than surveilling students or replacing the teacher's voice?
Students already live with AI. The guardrails haven't caught up.
The design assumption isn't "introduce AI into an AI-free classroom", it's that students already have ChatGPT open, and they're aware of the risks. The gap is between that awareness and their behavior.
Sources: College Board research brief on U.S. high-school GenAI use; project survey synthesis. Teachers show the mirror image, AI adoption is highest for behind-the-scenes work (64% to improve materials, 57% for feedback) and lowest for live, student-facing judgment (14% for 1-on-1 instruction). That hesitation is exactly where trust has to be earned.
Five methods: what students feel, what teachers fear, what the science says.
Built conviction by cross-checking primary voices against secondary evidence
6 student interviews
- Surfaced fear of judgment, curiosity that goes unmet mid-lesson, and repetition that kills motivation.
- Uncovered a striking behavior: students self-imposing guardrails ("you can't give me the answer") to protect their own learning.
- Revealed a shared instinct to develop their own ideas before AI shapes their thinking.
Survey synthesis
- Mapped where teachers actually use AI today vs. where its potential sits.
- Quantified student overreliance concern and school-level access policy.
- Grounded the whole project in a "students already have it" reality.
15-paper literature review
- Grounded design in established theory: Bloom's, Vygotsky's ZPD, UDL, Self-Determination Theory.
- Identified the Zimmerman self-reflection gap, students skip the phase that lets them improve over time.
- Confirmed human-in-the-loop beats full automation for trust and outcomes.
Competitive & discourse analysis
- Studied 11 tools from GoGuardian to Khanmigo, both models and cautionary tales.
- Analyzed Reddit, forums, and teacher op-eds for real classroom sentiment.
- Framed the north star: trust over detection, learning over "time on app."
In students' own words
"Before I ask AI anything, I like to develop my own ideas so AI doesn't shape my thinking too much."
"Sometimes I get lost in class and am too scared to ask the teacher, scared of judgment from other kids. So it's hard to raise my hand."
"I told the model that every time I want the answer, it can't give it to me. I have to show the steps until I get it."
"You learn so many things you never revisit. It'd be useful if it remembered what I wanted to remember and reminded me later."
"I used to love math, but now I find it a chore."
"Writing is an emotional process of getting your personal memory into the world. A model doesn't really know what matters to you, unless you tell it."
Competitive teardown
Each competitor showed what is doing well, and exactly what to avoid.
| Tool | What it taught me | Verdict |
|---|---|---|
| GoGuardian | Heavy monitoring and content-blocking keep students on-task briefly but manufacture a surveillance feel that damages trust. | Cautionary |
| Magic School | Powerful teacher tooling, but skews toward teacher convenience, material generation, over authentic student growth. | Cautionary |
| NotebookLM | Source-grounded answers dramatically cut hallucination, a strong model for the "resource-only" grounding mode. | Adopt |
| Khanmigo / SchoolAI | Socratic restraint: guide with questions, don't hand over answers. Districts trust it because it won't let students "practice wrong." | Adopt |
| DreamBox | Adapts to how a student solves a problem, not just right/wrong, a benchmark for real personalization. | Adopt |
| Packback | Students write their own questions (great), but its "Curiosity Score" can penalize creative answers (a warning about auto-scoring). | Adopt + guard |
| Colleague AI | Closest full analog: teacher-in-the-loop, visible AI transcripts as a trust mechanism. Validated the dual-goal direction. | Adopt |
Nine design principles.
Translating what I have learned to something that can be designed
Trust over surveillance
Design classrooms where AI use is transparent and learning is visible, instead of building detectors to catch misuse.
A 30 / 70 assist
AI carries the low-level load; the student owns the thinking, the decisions, and the final work. The thinking stays human.
Show the steps first
An incremental hint ladder, never the full answer, plus a guardrail that refuses "just solve it" and redirects to reasoning.
Protect the first draft
An opt-in mode keeps AI silent until the student commits to their own ideas, so it can't shape their thinking too early.
Grade the process
Capturing brainstorming, drafts, revisions, and reflection to show that learning happens along the way and not just the final work.
Make AI use visible
Disclosure is automatic and built-in; a private reflection log helps students see their own patterns over time.
A human, one tap away
A persistent path to a real teacher, visually distinct from any AI chat, so a person never feels like the fallback.
Keep the teacher in the loop
AI drafts narrative feedback (trusted); numeric scores always require teacher confirmation (not trusted). Agency stays human.
Design with student voice
The learner model is visible and editable to the student, personalization built with them, not for them.
Two interfaces, one system: a student experience and a teacher experience that mirror each other.
Both share a single cross-cutting assistance engine, assistance depth, grounding mode, "show steps first," and a "protect my first draft" toggle, so behavior is consistent everywhere AI appears, rather than reinvented per screen.
Two elements persist on every screen: a context-aware Ask AI button, and a Talk to Teacher icon in a distinct color, so a real person is always one tap away.
The incremental hint ladder
01 Protect My First Draft
An opt-in writing toggle that keeps the AI panel collapsed, no ideas, no interpretations, until the student marks their own first draft complete.
02 Projects graded as a process
A milestone timeline, planning memo, draft, feedback, revision, final, reflection, where every stage is saved permanently and forms a visible portfolio.
03 Judgment-free discussion
Students post their own questions with an anonymous-to-peers option, and a Creativity flag routes unconventional posts to a teacher instead of auto-scoring.
04 Reflection on My AI Use
A private journal of disclosure summaries and check-ins, with a gentle weekly prompt: "was there a moment you leaned on AI more than you wanted to?"
05 A student-visible learner profile
The mastery map and the insights the system notices ("you understand new concepts best through diagrams") are visible and editable by the student, with granular privacy controls.
06 Quiet by default
Three calm stat cards, tasks due, days to next test, mastery movement. No points, streaks, or leaderboards, and notifications collapse into one daily digest.
A persistent Preview as Student toggle lets a teacher see exactly what a student sees before changing anything, closing the visibility gap that the research identified as the single biggest disconnect.
01 Narrative feedback, not AI scores
The grading queue proposes narrative comments and a suggested score, but the number never reaches the gradebook without explicit teacher confirmation.
02 AI drafts, the teacher's voice lands
Pre-submission feedback can come from AI in the moment; post-submission feedback is authored by the teacher, with AI as an optional starting base, clearly labeled by source.
03 Supervised AI tutor
Teachers write custom instructions, preview any student's live conversation, and read condensed summaries instead of raw transcripts, oversight without surveillance.
04 Aggregate Growth Insights
Common questions, misconceptions, and progress surface at class and student level, turning everyday AI activity into a reteaching plan the teacher didn't have to build by hand.
05 A gentle variation nudge
If a class has run the same lesson format for weeks, the system suggests variations, privately, on the teacher's home only, worded as support, never as a performance flag.
06 Closing the operational gaps
Complementary-strength grouping, accommodation tagging, a shared library of what worked, approved parent summaries, gradebook sync, and a first-login walkthrough.
I also wrote the implementation guidebook.
Beyond the interfaces, I turned the research into a standalone handbook anyone can pick up and read, translating the findings into concrete, scenario-by-scenario guidance for exactly how the AI should behave, and when. It's written for three audiences at once.
Every recommendation is grounded in evidence, drawing on a 15-paper literature review, an 11-tool competitive analysis, and direct student interviews. It moves from background, to what the evidence shows, to the operational core: exactly what the AI should do, how, and when, for both the student and teacher experience.
- Introduction & purpose
- Foundational concepts
- What the evidence shows
- AI behavior, student experience
- AI behavior, teacher experience
- Still being developed
- Guidance for school & district leaders
- Guidance for policymakers & government
- Common pitfalls to avoid
- A note on student voice
- Glossary & sources consulted
Every feature traces back to a finding.
"Before I ask AI, I develop my own ideas so it doesn't shape my thinking."
The "Protect My First Draft" opt-in mode
Fear of asking a "bad question" in front of peers drives students to AI.
Anonymous-to-peers judgment-free question channel
"Just solve it" is where cognitive surrender begins.
Surrender-language guardrail + incremental hint ladder
Students skip the Zimmerman self-reflection phase, capping their growth.
Required reflection milestones + a "My AI Use" journal
Teachers value AI's narrative feedback but distrust its numeric scores.
AI proposes the score; the teacher must confirm it
Monitoring tools (GoGuardian) manufacture distrust.
Two-layer transparency; aggregate, non-surveillance insights
What I'd do next, and what I learned.
- Conduct a Co-design Interviews surfaced student voice; the next step is putting students and teachers in the room as co-designers of the actual screens.
- Get into a real classroom I'd spend a day observing how kids actually use these tools in situ, behavior in the wild always diverges from what people report.
- Interview teachers directly This round leaned on teacher-codesign literature; primary teacher interviews would sharpen R2 further.
- Define the fuzzy boundary "How much AI wording is too much" in a teacher's final feedback is still a judgment call, it needs a concrete, testable rule before it can ship.
The takeaway I'm proudest of
What I learned is that the strongest version of this product isn't how to make the most capable AI, but how to design the most honest one. Designing for restraint, transparency, and the student's own thinking turned out to be harder, and more interesting, than designing for capability.
It reframed the whole project from "what can AI do for learning" to "what should AI refuse to do, so learning still belongs to the learner."
This project is one of several where I take an ambiguous, high-stakes problem and carry it end-to-end: primary and secondary research, strategy, and a fully-specified interface. Happy to go deeper on any decision here.