The Caring Learning Agent: Open Learner Models, Metacognition and AI Assistance

MakingMaking the system's model of the learner visible, negotiable, and accountable

Any system that adapts to a learner has to respond to more than the learner's latest sentence. It needs an account of what the learner is trying to achieve, what they have attempted, which evidence they have produced, which concepts appear secure, where contradictions remain, and what help has already been offered. That account is a learner model. It may be no more than a record of topics completed, or it may combine activity histories, concept maps, confidence judgements, inferred misconceptions, goals, strategies, and predictions about future performance.

The existence of such a model raises a question of power. The system may make consequential claims about the learner while the learner cannot see how they have been represented, which evidence supports the representation, or how it shapes what the system does. An Open Learner Model changes that relationship by making the system's account available as an object the learner can inspect and act on. The learner can compare it with their own understanding, identify missing context, challenge a mistaken inference, and use the result to plan. The model stops operating only behind the interface and becomes a mediating tool within the learning activity.the model is a tool, not a verdict

The central question is therefore not only how an AI assistant can model a learner. It is how an assistant can construct and use a model of the learner while keeping that model visible, contestable, evidence-bearing, provisional, and oriented towards the learner's flourishing.

What "caring" means here. In this page, caring refers to attentive, adaptive, and accountable educational action: the assistant notices what the learner's evidence shows, adjusts its support to it, and makes its reasoning open to challenge. It is not a claim that the software experiences concern, feels warmth, or holds an attitude towards the learner. Where the word appears below, it should be read in that narrower, architectural sense.

1. The Learner Model and the Open Learner Model FoundationalKnowledge that endures for decades — core principles

TheThe two terms need separating. A learner model is a computational representation a system uses to make claims or predictions about a learner. It might record concepts encountered, answers given, errors and misconceptions inferred, assessed competence, strategies used, confidence judgements, goals, preferences, activity histories, requests for help, patterns of revision, and current and past plans.the model is the machine's internal state, not the UI

An Open Learner Model is a learner model, or a meaningful representation of it, made available to the learner and, where appropriate, to others who support learning. Susan Bull and Judy Kay describe such a model as making the machine's representation of the learner available to the learner, showing knowledge, misconceptions, and progress, and connecting that visibility with metacognitive activity and self-directed learning[1]. The opening changes the learner's role:

Hidden model:
  System interprets learner  ->  system adapts

Open model:
  System proposes interpretation
        ↕
  Learner inspects, challenges and revises
        ↕
  System and learner decide what to do next

An OLM is therefore more than a dashboard placed on top of analytics. It is a possible interface for shared metacognition and negotiated educational action.

2. Why an Assistant Needs a Learner Model Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AA general chatbot can answer the present prompt without keeping a durable model of its user. A learning assistant has a more demanding role. It may need to support activity across several sessions, multiple artefacts, changing goals, cumulative conceptual development, formative feedback, incomplete or contradictory evidence, periods of disengagement, shifts in confidence, and collaboration with teachers and peers.

To do this it has to keep apart at least what the learner has said, what they have attempted, what they have produced, what the evidence currently supports, what the system has inferred, what the learner believes about themselves, what remains uncertain, and what action has been agreed. These must not be merged. Consider a learner who says "I understand recursion." Recorded evidence shows that they correctly traced two recursive functions. The system may reasonably infer that they can trace simple linear recursion. It has no evidence yet about base-case design or mutual recursion, so the next step is to invite them to construct and explain a recursive solution.the model is a living hypothesis, not a label

The system should not compress all of this into "Recursion mastery: 78%". A score can serve a particular interface purpose, but it conceals the different evidential and ontological status of the claims beneath it.

3. The Architecture Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AnAn assistant that maintains an OLM can be described as a cycle of perception, interpretation, modelling, planning, action, and observation. Each step produces material the next step depends on, and the observation of what the learner does next feeds back into the model:

PERCEIVE     learner messages, artefacts, actions and context
    │
    ▼
INTERPRET    concepts, goals, difficulties and evidence
    │
    ▼
MODEL        domain, task, learner and self-models
    │
    ▼
PLAN         an educationally appropriate response
    │
    ▼
ACT          question, explain, retrieve, prompt, challenge or defer
    │
    ▼
OBSERVE      what does the learner do next?
    │
    └────────────► model revision

The OLM sits across the boundary between the system and the learner. The system's internal models feed a selected representation, with its explanation and provenance, into the OLM. The learner's own self-model sends back correction, interpretation, consent, and negotiation. The OLM is not identical to either model. It is an interface through which the system's representation and the learner's self-understanding meet.like a shared whiteboard in a meeting

4. The Classic Architecture, Extended Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe established architecture for intelligent tutoring systems distinguishes three components: a domain model (what is being taught), a learner or student model (who is being taught), and a pedagogical or tutoring model (how teaching proceeds). This three-part division is the conventional starting point for intelligent tutoring systems, and it is the one this page extends. For an agentic learning assistant the structure needs extending:like a map, a compass and a guide

WHAT         domain and curriculum model
WHO          open learner model
WHY          learner goals, curricular purpose, educational values
HOW          pedagogical strategy and scaffold selection
WITH WHAT    tools, sources, people and knowledge stores
UNDER WHAT   permissions, policies, consent and institutional rules
AUTHORITY
WITH WHAT    risk, wellbeing, agency and future opportunity
CONSEQUENCES

This extension prevents personalisation from becoming a purely technical match between a score and a resource. It forces the questions of purpose, authority, and consequence into the design.

5. "ITSs Care, Precisely" FoundationalKnowledge that endures for decades — core principles

Self'sSelf's phrase, "ITSs care, precisely", combines two commitments. The first is care in the sense that the system attends to differences among learners rather than delivering the same response regardless of evidence, context, or need. The second is precision: the educational grounds for adaptation must be represented explicitly enough to implement, inspect, test, and criticise[2]. The claim was not that the system has emotions or feels compassion.

That gives a demanding interpretation. A caring learning agent is not one that merely sounds warm. It is one that attends accurately and responsibly to the learner, remains alert to the limits of its own model, and acts in ways intended to support learning, agency, and wellbeing.

The distinction needs to stay visible. Affective performance says "I care about your progress." Educational care says "I have noticed this pattern, here is the evidence, here is my uncertainty, and here are several actions you may choose." Warm language may support interaction, but it is not evidence of care.not warmth, but evidence-based action

6. Care as an Architectural Property Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

CareCare should not be treated as an agent persona or a conversational style. An assistant can use reassuring language while maintaining an inaccurate learner model, hiding consequential inferences, over-scaffolding the learner, removing productive struggle, reporting sensitive information without consent, narrowing future opportunities, creating dependence, optimising assessment performance rather than learning, or treating institutional goals as if they were the learner's goals.productive struggle is not a bug

The reverse also holds. A system may communicate plainly and still be educationally caring, because it preserves learner agency, distinguishes evidence from inference, represents uncertainty, makes its model inspectable, invites correction, protects private information, notices when intervention is unwelcome, supports rather than replaces metacognition, declines to make unsupported identity claims, and escalates to human support when that is needed.

Care should be visible in the architecture of the relationship, not merely simulated in the language of the interface.

7. Levels of Openness Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

OpennessOpenness is not a single switch. Bodily and colleagues, reviewing open learner models and learning analytics dashboards, describe degrees of learner control including inspectable, cooperative, editable, persuasive, and negotiable forms[3]. Bodily and Verbert's review of student-facing dashboards and recommender systems covers the same field from the learner's side[4]. Read together, these suggest a sequence of increasing learner capability.

1. Inspectable. The learner can see the representation: confident explanation of concept A, inconsistent application of concept B, no current evidence for concept C. This supports awareness, but the learner remains largely a recipient of the system's interpretation.
2. Explainable. The learner can see why a conclusion was reached, such as the quiz from 2 October, draft version 3 of a project, and an explanation given in session 12.
3. Contestable. The learner can challenge an inference: "I used a different method because the assignment required it. This does not indicate that I cannot use the original method."
4. Negotiable. The system and the learner can hold competing interpretations. The system suspects difficulty selecting an algorithm; the learner believes the difficulty lay in an ambiguous requirement. The status is recorded as unresolved, and a contrasting task is proposed to gather further evidence.
5. Editable. The learner can correct appropriate parts of the model: clarify a goal, correct context, reject a purported preference, mark evidence as unrepresentative, add missing activity, request deletion, or revise an agreed plan. Editing should not allow a learner to replace assessed evidence, but it should let different claim types be represented clearly.
6. Agentive. The learner controls how the model is used: use it to recommend resources but not to predict final attainment; share some elements with a tutor; keep reflections private; revisit an inference after the next assessment.

The most important form of openness may be neither viewing nor editing. It is governance.

8. Evidence, Inference, and Identity Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe OLM should separate claim types visibly and semantically. An observed event ("the learner submitted three drafts") is different from performance evidence ("two drafts did not justify the selected method"), which is different again from a local inference ("the learner may need support connecting method choice to evidence") and a strategic proposal ("ask the learner to compare two methods against explicit criteria"). The statement "the learner is a weak critical thinker" is a far stronger and usually less defensible compression of all of these.

The further a claim moves from recorded activity towards capability or identity, the greater its evidential and ethical burden. The model should prefer phrasing such as "evidence currently supports", "evidence does not yet address", "the system has provisionally inferred", "the learner interprets this differently", and "further evidence could distinguish between". It should avoid "you are", "you always", "your learning style is", and "you lack ability in". Those forms state a conclusion about the person that the evidence does not carry.labels stick; learners internalise them

9. The OLM as a Metacognitive Tool Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AnAn OLM becomes educationally valuable when it helps a learner regulate their own activity. Bull and Kay connect opened learner models with metacognitive processes and self-directed learning[1], and the broader metacognition literature treats monitoring and control of one's own cognition as the central activity involved[5]. The OLM should support questions such as: what do I currently think I understand? What evidence supports that confidence? Where do my confidence and performance disagree? Which concepts remain disconnected? Which strategy have I repeatedly used, and what happened when I changed it? Where am I relying on a single source or a single form of evidence? What should I attempt next?

The cycle that supports this runs as follows: the learner acts, the system records evidence, the system proposes an interpretation, the learner inspects and responds, the model is revised or the disagreement retained, the learner chooses or negotiates the next action, and new evidence is produced. This is not only adaptive instruction. It is co-regulated model building.

10. The OLM and the Learner's Self-Model Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe learner's self-model and the system's learner model must not be merged. A learner may believe "I understand this, but I find timed assessments difficult". The system may hold "performance drops under time-constrained conditions". The institutional record says "Assessment 2 score: 48%". A tutor may observe "the learner's explanations are stronger in discussion than in writing". These can inform one another without becoming one authoritative description.

The OLM should therefore retain agreement, disagreement, uncertainty, contextual qualification, source-specific interpretations, and change over time. Take the claim "the learner avoids mathematical tasks". The evidence is that two optional activities were not attempted. The learner responds that they prioritised required project work because of limited time. The revised status becomes: insufficient evidence for avoidance, and the pattern may reflect workload prioritisation. The learner's response altered the operative model. That is the difference between an invitation to respond and a response that changes the outcome.

11. Learning Histories and Versioned Identity Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe model should preserve change rather than replace earlier states with the latest score. A learner might be recorded as able to identify a concept but not apply it at one point, as applying it with prompting later, as applying it independently in a familiar context later still, and as transferring it to a new domain and explaining its limits at the end. Development lies partly in those transitions. The OLM can offer timelines, concept histories, evidence trails, confidence changes, strategy changes, significant contradictions, learner annotations, unresolved questions, and abandoned plans.

A history should not be turned automatically into a polished story of linear progress. A learner may say that the system describes steady development while they experienced a period of confusion followed by one important conceptual change. Both accounts can remain visible. The model should keep enough history to support narrative construction without pretending to author the learner's definitive story. The same concern appears in the series page on learning logs and narrative identity.

12. Contradiction as a Shared Object Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AnAn OLM is a useful place to represent contradiction. Candidates include high confidence with weak explanation, strong coursework with weak timed performance, a declared goal with repeated avoidance, conceptual knowledge with poor transfer, an institutional classification that the learner disputes, a current capability that an outdated model fails to capture, and two sources of evidence pointing in different directions. The system should not resolve these automatically into one number. A contradiction can be presented with the candidate interpretations laid out side by side:

CONTRADICTION DETECTED

Learner confidence:   high
Current evidence:     successful familiar examples;
                      unsuccessful unfamiliar example

Possible interpretations:
  1. Knowledge is context-bound.
  2. The unfamiliar wording caused difficulty.
  3. The task required an unpractised strategy.
  4. The available evidence is insufficient.

Suggested next step:  choose a contrasting example and
                      explain the strategy before solving it

This turns contradiction into a resource for metacognition rather than a hidden trigger for remedial content. It also keeps the principle from the series page on productive struggle: not every contradiction belongs to the learner. The difficulty may arise from ambiguous teaching material, inaccessible design, an unsuitable assessment, missing prerequisite instruction, conflicting feedback, an incorrect domain model, or a mistaken system inference. The model must be able to represent each of these possibilities.

13. Productive Struggle and Intervention Restraint Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AA caring agent must decide when not to intervene as well as how to help. Premature assistance can remove the reasoning a learner needs to practise, reduce opportunities for diagnosing errors, reinforce dependency, confuse task completion with learning, prevent the learner from testing their own strategy, and conceal a useful contradiction. The learner model should therefore record the interaction as well as the knowledge state:

Current challenge:      debugging a recursion error
Evidence:               failing input identified; base case not traced
Productive activity:    constructing a trace by hand
Intervention threshold: do not give the correction yet
Permitted support:      ask what happens at the smallest input
Escalation:             offer a worked trace if the learner stays
                        blocked after an agreed interval, or asks
                        directly for help

A response consistent with this might be: "You appear to be making progress. Would you like a question, a hint, an example, or time to continue independently?" Care is compatible with restraint. The site's earlier discussion of error messages and useful struggle makes the same point from the side of feedback design[see the error-messages page].

14. A Multi-Agent Learning Assistant Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AnAn OLM does not require one conversational agent to perform every role. The roles can be separated, each with its own inputs and limits. Coordination among such roles is a standard concern of multi-agent design[see the multi-agent systems page], and the questions of human-in-the-loop design and graceful degradation are treated in the page on auditable agentic systems[see that page]. The following arrangement shows the roles conceptually; they need not be separate deployed models.

15. Trust and Provenance in the Learner Model Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

EveryEvery claim in the model should, where possible, carry its provenance. A claim that a learner can evaluate source reliability in familiar cases might record the activity it rests on, the source type (a learner-authored comparison), the process that produced the interpretation, a confidence level, its limitations (only two sources, both supplied by the course), any learner annotation (the learner used the checklist rather than an independent judgement), and a condition for review (reassess during an open-source research activity). This connects the OLM to the calculus of trust[see the calculus of trust].

The system should be able to answer several questions about any claim: what evidence supports it, whether the work was genuinely the learner's, whether the evidence applies in this context, whether it is current, whether it reflects independent performance, whether it is contradicted elsewhere, who or what produced the interpretation, and what follows if it is wrong. Trust should then govern use. A low-assurance inference can suggest a reflective question. A moderate-assurance inference can recommend optional practice. A higher-assurance inference can adjust instructional sequencing. A claim with very high consequences needs several forms of evidence and human review. A provisional inference may justify a question. It should not automatically justify restricting a learner's opportunities.

16. Privacy and the Right Not to Be Fully Modelled Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AA genuinely open model must include the possibility of limiting modelling. A learner should be able to ask what information is collected, why it is needed, how long it is kept, who can see it, which decisions use it, whether it can be corrected or deleted, whether they can participate without supplying it, whether a reflection can be kept from becoming an institutional record, and whether private metacognition can be separated from assessed evidence.

The system should also practise representational restraint. Not every available signal should become a learner attribute. A message sent at 03:00 is an observation, not evidence of poor time management. Repeated requests for reassurance are an observation, not evidence of general low confidence. A caring agent should not infer psychological or identity characteristics merely because they might improve prediction. The ability to model a learner does not establish the right to do so.

17. Care Without Surveillance Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

PersonalisationPersonalisation tends to increase data collection. Care requires enough knowledge of the learner to respond appropriately. But surveillance can undermine autonomy, trust, experimentation, willingness to admit uncertainty, private reflection, and the possibility that an old label stops following a learner once they have changed. The answer is not maximal data behind a friendly interface. The design should use data minimisation, local and purpose-limited models, clear retention periods, role-based access, separation of private and assessable records, learner-controlled visibility, provenance-bearing claims, deletion and correction mechanisms, explicit uncertainty, expiry of weak inferences, and human review of consequential decisions. A caring system should be capable of forgetting.

18. Caring About Learning and Caring About the Learner Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TwoTwo orientations need distinguishing. Caring about learning means optimising progress, mastery, completion, retention, transfer, and successful assessment. Caring about the learner also considers autonomy, dignity, wellbeing, privacy, belonging, motivation, intellectual agency, future participation, and the learner's right to refuse or reinterpret. These can conflict. Persistent reminders may improve completion while reducing autonomy. An emotionally persuasive interface may raise engagement while encouraging dependency. A detailed model may improve prediction while making the learner feel continuously judged. A caring agent has to weigh more than educational efficiency.

19. Caring Does Not Mean Agreeing Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AA caring agent should not simply affirm a learner's self-assessment. If a learner says "I fully understand this topic" but cannot explain or apply it, care may require a respectful challenge. If a learner says "I am incapable of doing this", the system should not answer with unsupported reassurance. It should help inspect the evidence and narrow the claim: the current record shows difficulty with two tasks involving this representation, which does not establish a general lack of capability, and it can compare the two attempts to find where the difficulty begins. Care involves responsiveness and respect, not frictionless agreement.

20. Curriculum Ethics and the V3 Model Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe OLM should not treat the curriculum as an unquestionable ground truth. Where a curriculum requires an outcome, the system should still be able to represent why the outcome matters, how it was selected, whether it remains current, which evidence counts, which alternative forms of competence are recognised, what ethical or political concerns apply, and whether the learner contests its relevance. The V3 model used on this site is the classic V model with a third stroke for values, and the curriculum-ethics page applies it to decisions about what should be taught[see the curriculum page]. The learner model can feed back the lived consequences of such a decision:

V3 curriculum judgement
        │
        ▼
Topic included and taught
        │
        ▼
Learner activity and evidence
        │
        ▼
Open learner model
        │
        ▼
Patterns of benefit, difficulty and exclusion
        │
        ▼
Curriculum review

Learners are not merely adapted to the curriculum. Evidence from learners can help revise it. That is a form of institutional reflexivity, and it depends on the OLM being inspectable by the people responsible for the curriculum as well as by the learner.

21. Design Principles Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe following fifteen principles summarise the argument as design requirements for a caring Open Learner Model:

1. Open. The learner can inspect the model in a meaningful form.
2. Evidential. Claims link to the activity or artefacts that support them.
3. Typed. Observation, inference, prediction, recommendation, and identity claim remain distinct.
4. Uncertain. The model represents confidence, missing evidence, and alternative interpretations.
5. Contestable. The learner can challenge claims and supply context.
6. Negotiable. Disagreement can remain visible rather than being forcibly resolved.
7. Revisable. New evidence can change the operative model.
8. Historical. The model preserves development without treating the latest score as the whole history.
9. Purpose-limited. Information is collected and used for stated educational reasons.
10. Learner-governed. The learner has meaningful control over visibility, use, and sharing.
11. Pedagogically restrained. The agent does not intervene merely because intervention is possible.
12. Curriculum-aware but not curriculum-subordinate. The model relates activity to intended learning while allowing those intentions to be critiqued.
13. Escalatable. The system recognises when human educational or pastoral judgement is required.
14. Forgetful where appropriate. Weak, outdated, or unnecessary inferences expire.
15. Caring precisely. Educational concern is enacted through accountable, inspectable, and revisable decisions.

22. Six Learner-Facing Views Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AA learner-facing OLM might offer six views. My goals shows what the learner is trying to achieve, who set the goal, and whether the learner can change or qualify it. My evidence shows what has been attempted and produced, which evidence is current, and what is missing. The system's current interpretation shows what the assistant infers, with its confidence and reasons. My interpretation lets the learner say where they agree, what context is missing, and which claims they contest. Contradictions and questions shows where evidence, confidence, goals, or interpretations conflict, and what might help understand the conflict. Possible next actions offers choices such as continuing independently, requesting a question or hint, viewing an example, attempting contrasting practice, revising the goal, consulting a person, or challenging a curriculum assumption.

The assistant should not always label one action as "recommended". It may instead explain the trade-offs between them.

23. The Wider System Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

                 CURRICULUM MODEL
            purposes and outcomes
                       │
                       ▼
DOMAIN MODEL ◄── OPEN LEARNER MODEL ──► LEARNER SELF-MODEL
concepts and     evidence, inference,   goals, confidence,
relationships    uncertainty, history   identity, strategy
       │                 │                      │
       └─────────────────┼──────────────────────┘
                         ▼
                PEDAGOGICAL DELIBERATION
          explain, question, wait, challenge,
           retrieve, practise, or escalate
                         │
                         ▼
                   CARING ACTION
                         │
                         ▼
              NEW ACTIVITY AND EVIDENCE
                         │
                         └────────► revision

Trust,Trust, consent, provenance, privacy, and learner agency should surround the whole system rather than sit inside one component. The earlier series page on pedagogy for networked and agentic learning sets out the broader teaching cycle this architecture supports. The page on knowledge bases and self-models covers the representational side of the same argument.

24. What This Page Does Not Claim Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

AnAn Open Learner Model is not an objective picture of the learner. More learner data does not automatically produce better support. Transparency alone does not create agency, and learner editing does not make a model accurate. Personalisation is not inherently caring, and warm language does not demonstrate care. An AI system does not experience concern. A learner model should not infer stable learning styles. Contradiction does not always indicate a learner misconception. The curriculum does not provide neutral ground truth, and a single mastery score does not adequately represent learning. An agent should not maximise progress at the expense of autonomy or wellbeing. The claims made here are qualified accordingly: the available evidence suggests, the system has provisionally inferred, the learner contests this claim, further evidence is needed, and this information should not be used for consequential decisions without review.

25. Conclusion FoundationalKnowledge that endures for decades — core principles

AA learning assistant that adapts without exposing its learner model asks the learner to trust a representation they cannot inspect. An Open Learner Model changes that relationship by making the system's account of the learner available as a shared, contestable object. This openness does not make the model neutral or complete. Every learner model compresses a changing person and history into selected evidence, categories, and predictions. The ethical task is not to perfect a definitive representation of the learner. It is to keep a model useful enough to support action, modest enough to represent uncertainty, open enough to be challenged, and provisional enough to change.

John Self described intelligent tutoring systems as caring, precisely[2]. Both parts of that phrase matter in an age of agentic AI. Care without precision can become sentiment, persuasion, or paternalism. Precision without care can become surveillance, classification, and control. A caring learning agent should therefore do more than sequence content. It should separate evidence from inference, help learners inspect their developing knowledge, preserve disagreement, protect private reflection, exercise restraint, and recognise when its own model or the curriculum may be wrong. The purpose of an Open Learner Model is not to tell the learner who they are. It is to give learner and system a shared, revisable account of where the evidence suggests the learner currently stands, and to support a better-informed choice about where to go next.

The model should not prescribe who the learner is. It should help the learner decide what to do next. The further a claim moves from activity towards identity, the greater its evidential and ethical burden. Care should be visible in the architecture of the relationship, not merely simulated in the language of the interface.

Related Topics

• Networked Learning Assistants and Personal Learning Networks — the extension of this page to assistants that share learner models with permission, and to the networks learners build around themselves.
• The Self-Model as a Learning Tool — the learner's own model, which an OLM is designed to meet rather than replace.
• Learning Logs and Narrative Identity — the evidence-constrained histories that an OLM can draw on.
• Contradiction and the Productive Struggle of Learning — the developmental role of contradiction, which the OLM represents as a shared object.
• Scaffolding and GenAI — the broader question of how AI assistance should scaffold without replacing the learner's own work.
• Modelling the Self: Object-Oriented Cognition Applied Reflexively — the architectural treatment of self-models that this page applies to learners.

References

1. S. Bull and J. Kay, "Open Learner Models," in R. Nkambou, J. Bourdeau and R. Mizoguchi (Eds.), Advances in Intelligent Tutoring Systems, Studies in Computational Intelligence vol. 308, Springer, 2010. https://doi.org/10.1007/978-3-642-14363-2_15
2. J. Self, "The defining characteristics of intelligent tutoring systems research: ITSs care, precisely," International Journal of Artificial Intelligence in Education, 10, 1999, pp. 350–364.
3. R. Bodily, J. Kay, V. Aleven, I. Jivet, D. Davis, F. Xhakaj and K. Verbert, "Open Learner Models and Learning Analytics Dashboards: A Systematic Review," Proceedings of the 8th International Conference on Learning Analytics and Knowledge (LAK '18), 2018.
4. R. Bodily and K. Verbert, "Review of Research on Student-Facing Learning Analytics Dashboards and Educational Recommender Systems," IEEE Transactions on Learning Technologies, 10(4), 2017, pp. 405–418.
5. J. H. Flavell, "Metacognition and Cognitive Monitoring: A New Area of Cognitive-Developmental Inquiry," American Psychologist, 34(10), 1979, pp. 906–911.