Stop When the Learner Can Continue
How an AI coach should decide when it has provided enough help.
How much help should an educational AI coach give to a learner?
The obvious answer is: enough to help. But enough is difficult to recognize. Too little assistance leaves the learner stalled. Too much completes intellectual work the learner could do and thus removes the practice through which capability grows.
Learning scientists sometimes call this the assistance dilemma: when should a tutor provide information, and when should it withhold information so that the learner must generate something for themselves?
Generative AI makes this question urgent for educators. AI often solves the learner’s problem before the learner has finished describing it. But an educational setting requires a different measure of success. Accuracy, safety, and transparency are essential. They do not by themselves produce a strong learning experience. An educational AI coach must also preserve the learner’s opportunity to think, try, and revise.
The unit of success is not the completed AI response. It is the learner’s next productive move.
A learner can be lost without knowing where
A natural first question is, “Where are you stuck?” Sometimes a learner can answer precisely: a term is unfamiliar, a calculation went wrong, or two ideas seem to contradict one another. That information is immensely useful.
But the question also makes a demanding assumption. It assumes that the learner can inspect their own understanding and locate the missing or mistaken piece. Often they cannot. A learner can know they are lost without knowing where they left the path.
In that moment, “I don’t know” is not a refusal to participate. It is evidence. The AI coach should not respond by opening the explanatory floodgates. It should become curious, or more precisely, it should demonstrate curiosity.
Treat each intervention as a hypothesis
Suppose a learner is trying to explain why exposure to nearby nature might affect how people respond during a conflict. The learner says, “I don’t know how these ideas connect.” Lux, RokSpark’s AI coach, could provide a polished account of attention restoration, self-regulation, stress, and aggression. The account might be accurate. It might also do the connecting for the learner.
A smaller intervention would test a possibility: “What tends to happen to patience and self-control when someone is mentally fatigued?” If the learner can answer, the missing relationship may come into view. If not, Lux might offer one piece: mental fatigue can make it harder to inhibit an impulsive response. Then it can return the question: “Given that, how might a restorative setting affect conflict?”
Suppose the learner responds, “Perhaps the setting helps restore some of the self-control needed to avoid an impulsive response.” The answer may be incomplete, but the learner has made the connection. Lux should resist polishing it and return the work to the learner.
Each move does two jobs. It helps the learner proceed, and it gives the coach better evidence about what the learner understands. The assistance is not only delivered. It is also tested.
A practical rhythm for assistance
Invite a diagnosis, but do not require one. Give the learner an opportunity to describe the difficulty without assuming they can.
Test one hypothesis. Ask for a prediction, a first step, a choice between two possibilities, or a connection to something familiar.
Offer the smallest targeted nudge. Provide a cue, distinction, example, or missing piece that addresses the likely obstacle.
Return the work. Ask the learner to use the new information rather than merely acknowledge it.
Increase the help only if needed. If the learner still cannot proceed, form another hypothesis and offer a little more support.
This rhythm turns an educational intention—keep the learner doing the thinking—into behavior a system can follow.
This is not a rigid script, nor is minimal help always better. Some situations require direct explanation. The point is to make assistance responsive rather than automatic. A good coach should not withhold help to manufacture frustration. It should preserve the work through which understanding develops.
A stopping rule for Lux
As we build Lux, this rhythm gives us a practical stopping rule. Once the learner can make the next judgment, attempt the next step, or explain the next relationship, Lux should stop adding information.
Give enough help to restart the learner’s thinking, then get out of the way.
Getting out of the way doesn’t mean abandoning the learner. Lux remains available, observes what happens next, and offers more support when the learner needs it. Restraint makes that support more attentive, not less generous.
An AI system may know the next six steps. If the learner is ready to take one of them, that is enough for now.
The moment the learner can continue is not the moment the coach has done too little. It is the moment the coach has done its job.
A question for readers: What is the smallest intervention that has helped one of your learners continue?
The Learning Environment publishes a Monday essay, and a Thursday field note about what advances in AI mean for how humans learn. Subscribe to join the inquiry.