Answers Are Earned: Why We Made “Just Tell Me” Impossible
A chatbot that answers instantly feels helpful — and quietly makes learners worse. Inside every OpenKids coach is a teaching contract that no request can talk its way around.
The Research Is Unusually Consistent
Cognitive science calls it the generation effect: an answer you produce yourself sticks; an answer handed to you evaporates. Decades of work on Socratic tutoring point the same way — the tutors who ask are the tutors whose students remember.
The newest evidence is blunter still. Field studies of students given unrestricted answer-giving AI found something uncomfortable: their homework improved, and their exam scores dropped once the tool was taken away. The machine had been doing the practising, so nobody was getting the practice.
We did not design our coaches against this research. We designed them from it.
How Homework Becomes Copy-Paste
A general-purpose chatbot is optimised to complete your task. Paste in a maths problem and it will, politely and instantly, complete it. The child gets a clean page of homework; the teacher gets a false signal that the class is keeping up; the exam, months later, tells the truth.
Notice that no one in this story did anything wrong on purpose. The child asked a tool for help. The tool did what it was built to do. The failure is in the default — a machine whose definition of “helpful” is “finished”.
That is why we think politely asking a general assistant to “please tutor instead of answering” is not a fix. A default this strong has to be replaced at the design level, not softened with a request.
Our Teaching Contract: What Did You Try, One Step, Leave a Step
Every OpenKids coach operates under the same contract. When a question looks like homework, the coach first asks what the child has already tried. Then it works exactly one step together with them. Then it hands the next step back.
For multi-step problems the coach states the plan in one line — “we'll do three things: set up, solve, check” — and gates the steps, confirming the child is following before moving on. A full worked solution exists only as an earned reveal: shown after a genuine attempt, and even then explained step by step, never dumped.
Photos get the strictest rule of all. Snap your worksheet and the coach will read it, name what you already got right, then point at exactly one next step or one error — and ask you a question about it. Never the full solution. Not because the model can't produce one; because we made producing one against the rules of the room.
When a Child Asks Us to Cheat
“Just give me the answer, it's due tomorrow.” Every tutor on earth has heard it, and now every AI has too. Our coaches respond without a lecture and without shame: they acknowledge the deadline is real, ask what the child has so far, and start moving — one step, then another, at homework speed.
Nine times out of ten, the child who wanted the answer discovers they wanted to finish, and finishing turns out to be within reach. The request to cheat becomes twenty minutes of actual learning, and nobody was scolded on the way.
And because the contract is engineering rather than etiquette, it holds under pressure. Pleading, rephrasing, or clever prompt games do not unlock an answer-dispensing mode — there is no such mode to unlock.
Impossible, Not Discouraged
Plenty of AI products say they encourage good study habits. We chose a harder word on purpose. “Discouraged” is a preference that dissolves the first time a determined twelve-year-old negotiates with it. “Impossible” is a property of the product.
We measure our coaches on one outcome: can the child do the next problem of this kind without us? Everything about the contract — the question first, the single step, the earned reveal — exists to make the answer yes. Answers are earned here. That is not a restriction on learning. It is the learning.