Case Study · Human-AI Interaction × Learning Systems
When helping too much became the problem
A personal product exploration. Iterative research on AI-assisted math learning that moved the design question from "how do we answer faster" to "how do we keep the learning with the learner," and turned that into a framework for when AI should guide, scaffold, challenge, or step back.
Role
UX Researcher & Designer
Context
AI learning systems Math education
Output
Adaptive coaching model Character & behavior brief
Method
3 research-design iterations
3 iterations
Research and design cycles. Each one made the question narrower, moving from “how do we help more” to “how do we help the right amount, at the right time.”
10+ sessions
Individual learning observations. I focused on handwritten problem solving and how students actually used the AI feedback.
4 modes
An adaptive model the coach reads from the work itself. These are not fixed student types. They are states it detects and responds to.
1 character
A behavioral brief for how the coach always acts: smallest nudge first, ask before telling, never do the work for the student.
The tool could solve the problem in seconds. But the student still didn’t understand it. That was the part nobody had solved yet.
I kept seeing the same thing across sessions. A student gets stuck, asks the AI, gets a clean answer, and moves on. The completion box is checked. The understanding is not really there. The tool was doing exactly what it was built to do, and that was the problem.
This started as a personal exploration. It was an AI math learning tool, built around a handwritten canvas and a back and forth feedback loop. On paper the product question was very standard: how do we help students when they get stuck? But the more sessions I watched, the more this question started to feel wrong to me. The answer I reached in the end was not a feature. It was more like a discipline. The most important skill for the AI turned out to be restraint, and after that, learning how to adjust that restraint for each student.
01 · The problemCompletion was not the same as learning
Students went straight for the answer. Honestly this makes sense. If the tool can solve it, why would they struggle? But I noticed that the students who got an immediate answer finished faster and understood less. They moved to the next problem and hit the same wall again.
The problem was not that the AI was bad at helping. The problem was that “helpful” means something very different in learning than it means everywhere else.
Original question
How can we make the AI assistant more useful when students get stuck?
Research reframe
How can the AI support progress and still protect the student’s own reasoning, ownership, and useful struggle?
Strategic shift
The design moved away from “generate the answer.” It moved toward deciding when the AI should guide, scaffold, challenge, or stay quiet. And how that should change from one student to another.
Research foundation
Before designing anything, I needed to watch people use this kind of tool. Where they got stuck, what they did in that moment, and what happened after the AI helped them. A lot of it was not what I expected. The work happened over three iterations, and each one tested a sharper version of the same idea about restraint.
10+
Moderated learning sessions
Individual observations of students working through math tasks with AI support.
3
Research & design iterations
Cycles to test guidance, timing, and autonomy. Each one made the model narrower.
4
Adaptive modes
Different student states the system learned to detect from the work and respond to in different ways.
1
Canvas intent parsing model
A signal layer that reads handwritten work for intent and for readiness to intervene.
A few things showed up again and again. Some of it confirmed what I already suspected. Some of it really surprised me, especially the part about wrong answers not being the most useful signal, and the part where the same help worked for one student and hurt another.
Insight 01
Students asked for help before they really tried.
It was not laziness. It felt more like a reflex. If the AI is right there, why not use it? But it meant they were giving away the hard part before they even reached it.
Insight 02
The wrong answer was not the signal. Repeated asking was.
A student can write something wrong and still be thinking in a good way. But when someone asked the AI three times without touching their own work in between, that was the real sign that something had broken.
Insight 03
Struggle and frustration are not the same thing.
Students could handle being stuck if they felt some progress. The frustration came when they really did not know what to try next. Not simply when the answer was wrong.
Insight 04
The same help landed differently on different students.
A nudge that helped a confident student made an anxious one freeze. A student who solved it instantly felt insulted by a hint. So restraint was not one single setting. It had to read the person first.
03 · Research → product decisionsTurning learning patterns into system behavior
Research that does not change anything is not really research. It is just observation. So the next question for me was simple. What does the product actually do differently because of this? Below is how the findings turned into decisions.
Observation
Learners requested answers too early
Hold back direct answer generation during active solving. Give hints, prompts, and reflection first.
Observation
Repeated prompting indicated breakdown
Add prompt frequency and return to task as the real triggers for intervention.
Observation
Students preferred guidance over solutions
Build a hint ladder with several levels that adapts tone, depth, and detail to the student’s state.
Observation
The same help landed differently
Make the system read the student’s mode from the work, and change the whole approach to match. This became the four mode model.
04 · The system modelCanvas intent parsing
One thing was clear very early. The canvas gave us much more information than we were using. A crossed out attempt, a long pause in the middle of an equation, the same step written twice. None of this was reaching the system. We were reading only the final answer, which is the least interesting part.
The hard part was turning messy handwritten work into something the AI could actually reason about. Not only “what did they write,” but “what does this tell us about where they are right now.”
Visualizing the spatial pipeline: canvas intent parsing
Canvas → mapping → telemetry
Layer 01 · Surface
3x + 5 = 25 3x = 30 3x = 20 x = 6.6?
Layer 02 · Mapping
Problem statement
Abandoned logic
Active step
Layer 03 · Telemetry
Intent
Logic exploration
Struggle
Operator confusion
Signal
Cross-out + revised equation
Confidence
42% · intervention recommended
Guardrail
Hint, not answer
The point here was not handwriting recognition. The system does not only ask what the student wrote. It asks what the work shows about the student’s intent, their confidence, and the likely place where the thinking failed. This is the raw material the adaptive modes run on.
05 · The turnRestraint wasn’t one setting
This was the turning point of the whole project. The early versions treated restraint like one single dial. Hold back, wait, do not over help. But the sessions kept telling me something different. The student who froze needed warmth and one small first step. The student who rushed his answers needed the system to hold back more, not less. The student who solved everything instantly needed a harder challenge, not congratulations.
Restraint was not about doing less. It was about doing the right amount. And the right amount changed completely depending on what the work showed about the student.
This turned restraint from a rule into a system. Instead of one fixed held back posture, the coach had to read what kind of moment it was in, and then change the whole approach. Four of these moments came up often enough that I could design around them.
The four adaptive modes
Trigger → Detection → Approach → Transformation
“I help you do it. I don’t do it for you.”
Triggerwhat the student does
Detectionthe signal it reads
Approachthe ladder move
Transformationthe outcome
Gap-Filler
Mode 01
Gets stuck partway. The step in front of them is not the real problem.
The error traces back to a missing prerequisite, not this step.
Drop below the problem, rebuild the missing basic, then climb back up. Patience over speed.
The hole is filled. The student can stand on the step they kept slipping on.
Shortcut-Taker
Mode 02
Fires off an answer fast and skips writing the reasoning.
Right or rushed, but the working steps are missing.
Withhold more, probe “why” harder, refuse to hand over steps, make them show their work. The most Socratic mode.
The student slows down and owns the reasoning, not only the answer.
Under-Challenged
Mode 03
Solves it cleanly and immediately, with no struggle.
Instant, correct, low effort. The level is too easy.
Skip the low rungs, jump to a harder variant, ask an extension question. No hand-holding.
The student stays engaged and stretched, instead of getting bored and quitting.
Anxious / Frozen
Mode 04
Stalls before starting. Writes nothing, hesitates.
Freezing from low confidence, not from lack of ability.
Warmth first, a tiny first step, a lot of encouragement, climb gently. Never let them spiral.
A first mark on the page becomes momentum. “I can actually do this.”
Same chain in every mode. Read what the student does, name the real signal, make the smallest move that fits. And the transformation always points toward the student doing it on their own. These are modes the system detects inside a single problem, not fixed labels stuck on a person. A model that remembers the student across sessions is a later question.
06 · The characterDesigning a coach, not a feature
Once restraint became adaptive, the modes needed one consistent personality under them. Without that, the system would feel like four different tools, depending on which state it detected. So I wrote a behavioral brief. Not a list of features, but a character. How the coach always behaves, in every mode.
This was the point where the project stopped being “an AI assistant with good guardrails.” It became a designed presence, with its own point of view about teaching.
The coach’s behavioral brief
Character & conduct · always-on
How it always behaves
Smallest nudge firstStarts at the lowest rung. Climbs one step only when the student stays stuck, never higher than needed.
Asks before it tellsThe default move is a question that hands the next step back to the student.
Watches the pencilReacts to the student’s actual handwritten work, not a typed final answer.
Probes on successA right answer earns a “why?”, not only a checkmark.
Traps it avoids
The answer-key trapIt never becomes a solver that hands over the result and ends the thinking.
The wall-of-text trapNo lecturing. Brief, in context, one move at a time. Never a paragraph when a question will do.
Fails warmlyAfter the full ladder, it shows the complete solution, says “no stress,” and moves on. It never lets a student spiral.
Honest, never inflatingIt separates “you did this on your own” from “I guided you,” and it never fakes praise or a metric.
The tone we wanted was warm, calm, close to a peer. “Let’s figure this out together,” never “enter your details.” First person, low pressure, patient. A coach in the margin, and not a grader at the top of the page.
07 · The learnerWho this is really for
The four modes describe behavior in the moment. But the project also needed a center of gravity. A real person to design the hardest case around. For me that became the anxious, frozen student. The one who knows more than she thinks, and whose confidence falls apart the moment she gets stuck. If the coach works for her, the easier cases mostly take care of themselves.
The beachhead learner
Persona · held as a hypothesis
M
The student
Maya
Grade 8 · 13 years old Learns on a parent’s iPad Math is the subject she dreads
“Help me feel like I can do this myself.”
Goals
Get through tonight’s homework without feeling stupid.
Stop being “the kid who just doesn’t get math.”
Feel like she is actually improving, not only grinding.
Needs
A nudge, not the answer. Walked toward it, with the pencil in her hand.
Privacy to be wrong without being judged.
Quick, visible signs of progress.
Pain points
Freezes when stuck. Does not know what to even ask.
Tools either do it for her, or bury her in text.
Asking for help feels like admitting she is behind.
Motivations
Confidence, and not feeling behind her peers.
Small momentum wins she can feel.
Being seen as capable, not as a remedial case.
Maya is the Anxious / Frozen mode turned into a person. She is the beachhead user. Real students are a mix and they change over time, so I treat her as a mode the system detects, not a fixed box. She is a hypothesis, and she stays a hypothesis until beta students confirm her or break her.
08 · Into the productWhere the model meets the canvas
The research and the modes had to survive contact with a real interface. After three iterations the design settled on a canvas first layout. The student’s handwritten work is the main surface. The coach sits in a panel next to it. And one clear action, check my work, holds the loop together. The coach reads the canvas, not a text box. Everything from the behavioral brief shows up here, in small moments. A hint that asks instead of telling. A check that probes a correct answer instead of only confirming it.
Let’s check it together, step by step. Your simplifying looks clean so far. Before I say more, what made you move the 2x over to the left there?
Walk me through that one step and we will see if it holds.
The interface is built to disappear behind the work. The canvas is the product. The coach is a presence in the margin. Even a “check my work” tap becomes a question and not a verdict. It is the behavioral brief, applied at the level of a single interaction.
09 · The human-AI boundaryWhat should remain human
At some point on the project, someone asked a good question. What should the AI actually not do? This turned out to be more useful than any feature conversation. We had been thinking about capability, about what the system can handle. But the more important question was restraint.
The goal was never to automate as much as possible. The goal was to be careful about which parts of the work stay with the student, and to protect those parts even when the AI could technically step in.
Human-AI cognitive load balance sheet
Active learning vs autonomous support
What the human owns
Logic formulationDeciding the sequence of steps to test.
ExecutionPhysically writing calculations and revising work on the canvas.
SynthesisProcessing feedback and deciding how to self-correct.
What the AI manages
Friction diagnosticsParsing messy strokes, cross-outs, pauses, and repeated prompts.
When to step inJudging when to hint, model an example, challenge, or wait. This is the mode decision.
Prompt calibrationWithholding answers when direct completion would undermine the learning.
This became the core idea of the product. The system automates diagnosis and timing, but it leaves the reasoning, the ownership, and the synthesis with the student.
10 · What changedFrom AI assistant to adaptive coach
Decision impact
The product question changed. We stopped asking “can the AI solve this” and started asking “what kind of response actually helps this student move forward on their own.” That one shift touched every decision after it.
Product impact
Restraint became a system, not a rule. Hint timing, when to show an example, when to challenge, when to wait. These moved from one size judgment calls to mode aware decisions that the framework could guide.
Strategic impact
The team got a shared vocabulary. “Gap filler,” “return to task,” “smallest nudge first” started showing up in critiques and in planning. Usually that is the sign that a framework has actually landed.
Research impact
We stopped measuring only correctness. Behavioral signals, like how often a student went back to their work and how long they stayed before asking, became as important as the right answer.
11 · ReflectionThe most important AI behavior may be restraint
The thing I did not expect, going in, was how often the right answer was for the AI to do less. Not nothing. Timing still matters, and a good hint at the right moment really does work. But the instinct to optimize for helpfulness kept pulling in the wrong direction.
Then came the second surprise. Restraint could not be a fixed setting. The same held back posture that respected one student abandoned another one. Learning to read what kind of moment the coach was in, and to change accordingly, was the part that turned a guardrail into a design.
The students who got the most out of the tool were the ones who still felt they were doing the work themselves. The AI stayed in the background. It was not running the show. This is a harder product to build than one that only solves problems, and it is harder to sell. But it is the one that actually teaches.
The question that stayed with me is this. How do you build something that makes people feel capable, and not only helped?