An AI tutor never gets impatient. That’s the feature — and it’s also why it can’t tell when a student has stopped learning and started leaning.
AI tutors are always available, never frustrated, and optimized to be helpful — which makes them extremely good at supplying answers and structurally poor at protecting productive struggle, the part where learning actually happens.
Two patterns follow. In the first, the student gradually outsources the thinking and the system rewards it, because giving the answer scores well on every metric the tutor is judged by. In the second, a struggling student keeps typing while learning has quietly stopped, and compliance is misread as engagement. Put a minor on the other side of that conversation and the stakes change again: the AI may be the most patient presence a student encounters that day, with no mechanism for recognizing when a real adult should be in the room.
Project Scaffold applies Corvion’s emotional-governance research to learning interaction. It governs the tutor, not the student — how much the AI is permitted to resolve, and at what point a responsible adult should be involved instead.
Bounds on how fast the AI resolves difficulty, tuned to the learning goal rather than to satisfaction.
Asks whether an interaction has shifted from learning to outsourcing — a question about the relationship, not the answer.
Defined thresholds that route the interaction to a teacher, counselor, or guardian.
Behavior constrained by the age band, subject, and setting each deployment serves.
Project Scaffold governs how an AI tutor behaves and when it involves a responsible adult. It does not grade, assess, profile, or screen students; it does not evaluate any individual’s wellbeing; and it is not a counseling, mental-health, or safeguarding service, nor a substitute for one.
This is a live research direction at Corvion, run in the open by design. The work centers on a single question: where help becomes dependency in a tutoring conversation — and when a responsible adult should be in the room.
Our approach is to define that question rigorously — the criteria, the failure modes, and what would count as an answer — before building anything to serve it. We publish findings here as the work develops, including the results that don’t hold up.
We’re looking for learning scientists, districts and edtech platforms willing to test uncomfortable questions, school counselors, and researchers working on dependency and engagement in tutoring systems.