Internal Working Models
Internal working models are the mental representations of relationships that each person carries forward from early attachment experiences. Constructed through thousands of caregiver interactions across the first three years of life, they encode expectations about what to expect when one reaches out for connection. The models are not beliefs one could revise by argument; they operate at the level of perception itself, shaping what the person notices, remembers, and feels before conscious thought begins. Bowlby named them 'working models' because they function instrumentally — guiding behavior in ways the person does not perceive as guided. Their structural identity with Segal's fishbowl is not coincidental: both describe invisible assumptions that feel like reality because the person has never been outside them.
In the AI Story
Bowlby drew the concept from cognitive psychology's models of mental representation, combining it with psychoanalytic insights about transference. The resulting framework specified that early relational experience does not merely leave memories — it installs processing architectures that shape how all subsequent relationships are perceived and navigated. The person whose working model says 'reaching out is dangerous' does not consciously think this when someone offers help. She feels something closer to wariness, and the wariness organizes her behavior before reflection can intervene.
The parallel to Segal's fishbowl provides a conceptual bridge. Both frameworks describe structures that: (1) are formed through experience, (2) operate below conscious awareness, (3) shape perception rather than content, and (4) resist revision because they constitute the lens through which all evidence is processed. The orange pill moment — Segal's name for the crisis of visibility when assumptions suddenly become visible as assumptions — is, in Bowlby's framework, the activation of an internal working model under conditions that disrupt it enough to make it perceivable.
Revision of working models is possible but requires specific conditions. Bowlby insisted on two: the model must become visible (extraordinarily difficult because it is the lens, not the object), and the revision must occur within a relationship that provides sufficient security to tolerate the disorientation. This is what psychotherapy accomplishes — a relationship in which old expectations are activated and carefully, repeatedly contradicted by the therapist's actual behavior, producing the new model alongside the old that earned security research later documented.
The AI transition activates working models at civilizational scale. The compulsively self-reliant professional reads AI tools through a model that says 'help is unreliable, only self-sufficiency is safe' — and the tool confirms the model at every turn. The anxiously attached worker reads the same tool through a model that says 'my worth is contingent on performance' — and experiences the tool as a threat. Same technology, fundamentally different psychological response, governed by structures installed before any of these professionals knew the word 'intelligence.'
Origin
Bowlby introduced the term in the first volume of Attachment and Loss (1969) and elaborated it across the trilogy. The concept drew on Kenneth Craik's 1943 work on mental models and was refined through dialogue with cognitive psychologists including Inge Bretherton and Everett Waters.
The Adult Attachment Interview, developed by Mary Main in the 1980s, provided the empirical methodology for identifying working models in adults through the analysis of how they talk about their childhood relationships — not what they say but how they structure the telling.
Key Ideas
Operate below awareness. Working models are not beliefs but perceptual frameworks; the person does not experience them as assumptions but as reality itself.
Installed through experience. The models are built from thousands of interactions with caregivers in the first three years, before language and conscious evaluation become available.
Shape perception, not content. The model does not tell the person what to think; it determines what she perceives as threatening, rewarding, trustworthy, or dangerous.
Resist revision. Because the model is the lens, not the object, evidence that would contradict it is typically processed through it and therefore fails to challenge it.
Revisable in relationship. The only known mechanism for genuine working-model revision is a sustained relational experience that disconfirms the old model — what earned security research documents as achievable across the lifespan.
Debates & Critiques
Contemporary attachment researchers debate whether the working-model concept should be understood as a unified structure or as multiple domain-specific models. Mary Main's research suggested a unified organizational pattern; subsequent work by Kim Bartholomew and others has documented more fragmented patterns in which a person might hold secure working models for friendship while carrying insecure models for romantic relationships. The implications for AI — which may activate different working models depending on whether the user frames the tool as collaborator, competitor, or instrument — remain an open research area.
Appears in the Orange Pill Cycle
Further reading
- John Bowlby, Attachment and Loss, Vol. 1: Attachment (Basic Books, 1969)
- Inge Bretherton, 'The Origins of Attachment Theory: John Bowlby and Mary Ainsworth' (Developmental Psychology, 1992)
- Mary Main, 'The Adult Attachment Interview' (unpublished protocol, 1985–)
- Peter Fonagy et al., Affect Regulation, Mentalization, and the Development of the Self (Other Press, 2002)