The Self-Model as a Learning Tool: Metacognition, Regulation and Agency

PartPart 2 of Learning as Connection, Contradiction, and Self-Construction

Learning changes more than what a person can recall. It changes the strategies through which they take part in activity, and it changes the model through which they interpret their own capabilities. A student who has worked through a proof does not only hold the proof. They have also formed a judgement about whether they can do proofs, how long the work takes them, and what to try when it stops making sense. That judgement is a self-model, and it does work. It decides what the learner attempts next, when they ask for help, and whether a setback feels like information or like a verdict.

This article argues that metacognition is best understood as the use of a provisional, revisable self-model to monitor, interpret, and regulate learning activity. The self-model is useful because it is selective, and it becomes harmful when a local judgement hardens into a claim about identity. The aim of teaching is therefore not to make learners perfectly self-transparent. It is to make the self-models that matter sufficiently inspectable and revisable that they can be tested against evidence.logs make the model inspectable

1. What Learners Come to Believe About Themselves FoundationalKnowledge that endures for decades — core principles

OverOver a course of study a learner forms beliefs about what they know and do not know, what they can currently do, which strategies work for them, what they find difficult, when they need help, which sources they trust, whether they have actually understood something, and what kind of learner they take themselves to be. John Flavell's account of metacognition, which introduced the term into psychology, treats this kind of knowledge about one's own cognition as a central object of study, together with the monitoring of ongoing cognitive activity[1]. The beliefs above form a practical self-model: a working account of the learner that is used to choose what to do next.

2. Three Levels of Learning Activity Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

ItIt helps to separate three levels of activity that run together in practice. At the first level is object-level work: reading the article, writing the program, analysing the data, constructing the argument, testing the model. At the second level is metacognitive monitoring, the running commentary a learner keeps on that work:

I understand the definition but not the example.
I cannot yet explain why this method is appropriate.
This source conflicts with my interpretation.
I am repeating the same error.
My confidence exceeds the quality of my explanation.

At the third level is metacognitive control, the choice of what to do in response: draw a concept map, find a contrasting example, ask for feedback, try a smaller case, return to the evidence, change strategy, or stop and resume later. The three levels form a loop:

Learning activity
       │
       ▼
Evidence about performance
       │
       ▼
Self-model
       │
       ▼
Strategy selection
       │
       ▼
Revised learning activity

Most of the difficulty in self-regulated learning lies in the middle of this loop. A learner may collect evidence accurately and still interpret it through a self-model that was formed long before the current task, and so choose a strategy that fits the old model rather than the present situation.the self-model is a story you tell yourself

3. The Self-Model Is Selective FoundationalKnowledge that endures for decades — core principles

AA self-model is not a complete representation of the learner. It compresses experience into claims that bear on future action: "I understand recursion", "I find notation difficult", "I need examples before abstraction", "I work better under pressure", "I am not good at academic writing". Some of these claims are locally useful. Others are generalisations drawn from a narrow set of occasions.

Several questions help to sort them. What evidence supports the claim? In which contexts does it apply? At what level of specificity is it stated? Does it help select productive action? Can it be revised? And, the question that matters most in practice, has a temporary difficulty become a judgement about identity?

The escalation from one to the other is often quick, and each step is a larger compression:

Event:
  I did not understand this explanation.

Capability judgement:
  I do not understand this topic.

Identity claim:
  I am not the kind of person who can understand this subject.

The event is specific and bounded. The capability judgement generalises it to a topic, and the identity claim generalises it to the learner. Each step makes the claim harder to test, and so harder to revise.the trap of 'I am bad at X'

4. Opaque and Transparent Self-Models Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe Modelling the Self series distinguishes two ways a self-model can operate. An opaque self-model is a labelled entry that reasoning can consult and question. A transparent self-model feeds directly back into perception, so that nothing downstream can inspect it as a model at all; it simply shapes how a situation appears. The follow-up page examines that distinction further, but the same contrast is useful for learning without any commitment to the architecture behind it.

In a learner, an opaque self-model sounds like this:

I presently think of myself as poor at mathematics,
but that belief may rest on a narrow set of experiences.

A transparent one sounds like this:

I am poor at mathematics.

The second version does not present itself as a belief that might be wrong. It presents itself as a fact about the world, and so it governs what the learner notices and what they attempt. A transparent negative self-model can lead a learner to avoid difficult tasks, to read confusion as incapacity, to discount improvement, to treat asking for help as confirmation of inadequacy, and to withdraw before the feedback that would have revised the judgement arrives.

The aim is not complete introspective transparency, which is neither achievable nor necessary. It is to make the consequential self-models inspectable enough that they can be tested. Bjork's work on desirable difficulties adds a caution here. Conditions that make training harder can reduce performance during the activity while improving later retention, so a learner who judges their ability by how smoothly practice is going may draw the wrong conclusion from a session that was in fact doing useful work[2].smooth practice often means shallow learning

5. Self-Models Are Tools, Not Verdicts FoundationalKnowledge that endures for decades — core principles

AA pedagogically useful self-model is a tool for selecting action, not a verdict on identity. This distinction should govern how learning environments represent learners. Student analytics often produce labels such as "at-risk student", "low-engagement learner", "weak writer", or "high performer". These labels may begin as provisional readings of limited records. Once they are displayed, repeated, or used to decide which opportunities a student receives, they become part of the learner's environment and of how the learner understands themselves. The system can then help construct the person it claims only to describe.labels become the story the learner tells themselves

The same risk applies when a learner's own self-model is formed from a small number of records. Used as a tool, the self-model suggests what to try. Used as a verdict, it decides in advance what the learner is capable of.

6. Changing the Activity Rather Than the Judgement Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe Unstuck Ladder, described in the project-guidance pages, is a practical example of this kind of regulation. Its starting observation is that "I'm stuck" rarely means that all the work is stuck at once. It usually means that one kind of work, such as reading or writing, has stopped moving, while others, such as building or reorganising existing material, remain available. The ladder then moves the learner through a sequence of different kinds of work instead of pushing harder at the one that is blocked. The guidance on getting unstuck covers the same situation from the point of view of a final-year project.

In metacognitive terms the shift can be written as a change in the self-model's content:

Global self-judgement:
  I cannot make progress.

Revised activity model:
  This particular form of work is blocked.

Control response:
  Switch to a connected activity that remains available.

The first version produces a general conclusion about the learner and no obvious next step. The second describes a specific, local condition and points to a response. Nothing about the learner's ability has been asserted or denied. Only the scope of the problem has been stated accurately, and that accuracy is what makes the response possible.

Uncertainty works the same way. Treating an unresolved problem as a state of the task rather than a defect in the learner allows the learner to keep working on it. The navigating uncertainty page develops this idea further.uncertainty is a state of the task, not you

7. AI and Metacognitive Outsourcing Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AnAn AI assistant can support metacognition in several ways. It can help a learner identify gaps in an explanation, compare their work with stated criteria, suggest alternative strategies, retrieve earlier reflections, detect recurring difficulties, generate questions that test understanding, and connect present work with earlier activity. The scaffolding and GenAI page discusses how these uses depend on the learner still doing the work that matters.

It can also weaken metacognition. An assistant that supplies an answer before the learner has diagnosed the difficulty removes the diagnosis. An assistant that turns uncertainty into polished prose hides the uncertainty the learner needed to notice. An assistant that generates retrospective rationalisations supplies a reason for a decision after the fact, not the reasoning at the time. Fluent output can be mistaken for understanding, the assistant's model of the learner can replace the learner's own, and a conversation history can become the only record of the learner's activity.the danger of skipping the hard work of thinking

The design goal that follows is metacognitive augmentation, not metacognitive substitution. A system that helps the learner monitor and regulate their own activity leaves the self-model in the learner's hands. A system that performs the monitoring and regulation for them takes it away, and the learner's model of themselves then rests on the system's account.

8. What This Asks of Teaching FoundationalKnowledge that endures for decades — core principles

IfIf a self-model is a provisional tool that governs action, teaching has to do three things. It has to help learners notice which self-claims are operating on their decisions. It has to supply evidence that can test those claims, including evidence that appears only after a difficult session. And it has to keep the step from event to identity visible, so that a learner can see when a single occasion is being treated as a permanent fact.logs as the evidence trail

Metacognition is therefore partly a calculus of trust applied to oneself: deciding when to rely on memory, on present confidence, on a current strategy, or on a settled picture of one's own capacities, and when to check. The learner who can make that decision is not one who never doubts themselves. It is one who can say what they would need to see before revising a belief about themselves.

Related Topics

• When Meaning Becomes Architecture — the opaque and transparent distinction for knowledge bases and self-models in agentic systems.
• The Unstuck Ladder — switching between kinds of work when one stops moving.
• Scaffolding, the Zone of Proximal Development, and Using GenAI Well — the conditions under which assistance supports the learner's own work.
• Navigating Uncertainty — treating unresolved problems as properties of the task.

References

1. J. H. Flavell, "Metacognition and cognitive monitoring: A new area of cognitive-developmental inquiry," American Psychologist, 34(10), 1979, pp. 906–911.
2. R. A. Bjork, "Memory and metamemory considerations in the training of human beings," in J. Metcalfe and A. Shimamura (eds.), Metacognition: Knowing about Knowing, MIT Press, 1994.