The Cost of Productive Friction

AI tutoring creates a design problem that cannot be solved by maximizing either speed or persistence. This chapter compares what happens when a platform forces students to keep working, when an AI tutor slows recovery in order to guide reasoning, and when human support is used to sustain participation. The goal is a practical judgment: which kind of work is worth the time it consumes, and what evidence would show that the extra effort became learning rather than merely progress inside the platform?

A streak is a threshold, not a learning result

The NUMI experiment makes the distinction between platform progress and durable understanding unusually clear. Its mastery condition required students to watch at least one minute of an instructional video and answer three questions in a row correctly before advancing. If a student made a mistake, the streak reset. Students in the non-mastery condition could move through the activity more freely after attempting questions.

The rule changed behavior substantially. Among students assigned to CAL-only support, mastery increased the number of Exercise 1 questions completed by 4.62, the number answered correctly by 1.23, and the probability of reaching three correct answers in a row by 28.7 percentage points. It also increased time spent in the exercise by 6.69 minutes. Those are not trivial changes. The rule made students stay with the material, generated more attempts, and made the platform’s definition of success more common.

But the rule’s effect on the delayed assessment was close to zero. One week later, mastery alone did not clearly improve the combined practiced-minus-unpracticed score, the Exercise 1 contrast, the Exercise 2 contrast, or the total practiced score. If the three-answer streak were a reliable measure of durable understanding, its large effect during practice should have appeared as a comparably clear retention effect. It did not.

That result does not show that mastery learning as a broad educational approach is ineffective. It shows that this particular threshold was an imperfect proxy for learning that persisted beyond the interface. A student can reach three correct answers through genuine understanding. Another can reach the same streak after repeated exposure to similar items, a lucky sequence, or trial and error that produces a short-run success without a transferable concept. The platform records the same endpoint in each case. The delayed test asks whether the endpoint meant more than the ability to complete the immediate sequence.

The distinction matters because a progression rule has two possible functions. It can protect a learner from rushing past material they have not understood. It can also manufacture a visible completion signal by keeping them on a narrow set of questions until the required pattern appears. The rule creates persistence, but persistence becomes educationally productive only if the support during that persistence helps the student explain, retrieve, and apply the idea. More attempts are an opportunity for learning, not proof that learning occurred.

This is where AI and mastery may complement one another. Mastery makes a mistake consequential: the student cannot simply move on after a wrong answer. AI can use the resulting moment to ask a smaller question, address a misconception, and guide another attempt. Without the progression rule, a student may have little reason to engage with a tutor after an error.