Domain · Educational Technology

Mastery tracking.
Without understanding mastery.

AI tutors track correct answers, hint usage, and completion rates. They miss the 45-second hesitation that separates a student who is thinking carefully from one who has stopped understanding. VanLehn showed this gap in 2011. It remains accurate today.

AI tutors are good. Human tutors are better. This is why.

VanLehn's 2011 meta-analysis found that intelligent tutoring systems outperform conventional instruction but remain less effective than human tutors. The effect size gap is approximately 0.4 standard deviations. That gap has not closed significantly since the paper was published.

The reason is not processing power or curriculum quality. Human tutors read cognitive and emotional states in real time - they can tell when a student is confused but too embarrassed to ask, when they are bored and need a harder problem, when they are anxious and need encouragement before content. AI tutors cannot. They see the answer, not the state that produced it.

"The gap between AI tutoring and human tutoring is primarily a gap in psychological understanding - not in instructional content."
VanLehn, K. (2011). The relative effectiveness of human tutoring, intelligent tutoring systems, and other tutoring systems. Educational Psychologist, 46(4), 197–221.

ABSD proposes a framework for closing this gap - not by adding sensors or cameras, but by interpreting the interaction data that AI tutoring systems already collect.

EdTech systems reach Layer 2. Layer 3 is the frontier.

EdTech is ahead of most CX platforms on the ABSD ladder. Mastery models track behavioural patterns - attempts, hint usage, time-on-task - and use them to calibrate difficulty. That is genuine Layer 2 capability. But the cognitive-emotional state beneath the behaviour is invisible. The 45-second pause, the repeated failure, the avoidance of worked examples - all logged but not interpreted as psychological signals.

Action
Strong
Behaviour
Partial
State
Developing
Drive
Minimal

Based on VanLehn (2011), Li et al. (2020), Paquette et al. (2014). Indicative of sector average.

Seven signals in an EdTech context.

Hesitation
Student pauses 18 seconds before answering a question they have attempted before.
What to do: Offer a process hint - how to approach this type of question - not the answer. Preserves agency while reducing the cognitive load causing the pause.
Thrashing
Learner switches between lesson video, transcript, quiz, and notes panel without progressing.
What to do: A "where would you like to start?" prompt that simplifies the interface temporarily. Reduces the choice paralysis without removing options permanently.
Regression
Student returns to a completed chapter mid-exam preparation.
What to do: Surface the specific section most relevant to their current position in the curriculum. Do not send them back to the beginning of the chapter.
Acceleration
New student completes a module in 40% of expected time then fails the assessment.
What to do: A brief diagnostic before the next assessment. Surface the gap before they discover it in the test - better for confidence and for learning outcomes.
Persistence
Student attempts a coding problem six times over 40 minutes, making incremental progress.
What to do: Nothing. This is flow-adjacent behaviour. Systems that interrupt flow to offer help actively harm learning outcomes. Pekrun's research is explicit on this point.
Avoidance
Student consistently skips worked examples, going directly to practice questions and abandoning them.
What to do: Make worked examples feel like a choice rather than a prerequisite. "Want a hint?" activates competence drive rather than threatening it.
Flow
Learner completes three to four units per session, rarely seeks hints, returns consistently.
What to do: Gently increase challenge difficulty. A slight upward adjustment keeps the learner in flow. Too much challenge creates anxiety. Too little creates boredom.

If EdTech AI reached Layer 3 - what changes.

The mastery model stops confusing performance with understanding. A student who gets the right answer after 45 seconds of hesitation and one who answers in 4 seconds both show mastery at Layer 1. At Layer 3, the hesitation is a signal - the first student understood but needed cognitive processing time, the second answered reflexively. One is consolidating, one is recalling. They need different next steps.

Anxiety becomes visible before it becomes failure. Avoidance patterns are the earliest detectable signal of learning anxiety. A student who consistently skips worked examples is showing anxiety at Layer 2. At Layer 3, that anxiety is named. The intervention changes from "you haven\'t completed this section" to "here is a lower-stakes way in."

Flow becomes the optimisation target, not mastery. Pekrun et al. (2017) confirmed that sustained engagement produces better long-term outcomes than maximised short-term performance. A system that detects flow and adjusts challenge to maintain it - rather than maximising difficulty at every step - produces better learners over a semester, not just better test scores this week.

Competence drive shapes how the system scaffolds, not just what it teaches. Two students both reach a frustration state. One persists and one disengages. The drive beneath the state determines the correct intervention. SDT provides the theoretical foundation. ABSD provides the architecture to apply it in production.

See it traced
Watch a student hesitation signal travel from pause event to fear of failure.
Trace an EdTech signal → Take the survey