A child reaches for a caregiver after a frightening sound and receives help. Repeated experiences like this can inform what the child expects when help is needed later. Attachment theory calls such organized expectations internal working models. The phrase is a theoretical shorthand for representations of self, others and relationships that guide attention, interpretation and action. It does not mean a fixed script stored in one brain location. Nor does one comforting or disappointing episode create a permanent model. The question is how experience is represented, used and revised across relationships and development.
A person might expect one friend to respond reliably while feeling uncertain with a different friend. That is compatible with internal working models being partly relationship-specific. A model can also contain broader regularities learned across experiences. Treating it as one immutable 'attachment type' hides this layered organization. Researchers study representations through interviews, narratives and questionnaires, each of which captures a different slice of the construct. The value of the idea lies in explaining continuity without insisting that early experience determines every later relationship.
From interaction to representation
Early attachment theory proposed that experiences with caregivers help children form expectations about whether support is available and whether they are worthy of care. These expectations could affect how a child interprets distress and seeks comfort. That proposal connects observable behavior with an inferred representation. The inference requires care: a child who does not approach a caregiver in one observation may be tired, unfamiliar with the setting or responding to a local interaction. A working model is not directly visible from a single act.
Researchers use repeated observation and multiple measures to ask whether early interaction predicts later representation. Steele and colleagues studied a longitudinal cohort and measured young adults' secure base script knowledge through narratives about seeking and receiving effective support. That knowledge was associated with early attachment and observed caregiving sensitivity. The result fits the idea that relationship experience becomes portable knowledge. It does not show that a particular infant episode mechanically causes one adult story or that all individuals follow the same path.
The secure base script is one operationalization of an internal working model. A narrative may include a problem, a bid for help, a responsive other and a return to exploration or task. Researchers score how coherently such elements are organized rather than asking whether the participant recites a correct phrase. The representation is inferred from performance on a specific assessment, not seen directly. Someone may know the script while a current relationship remains unsafe. Conversely, a brief narrative can be affected by language, mood and the task.
A working model is also distinct from autobiographical memory. A person can recall an event accurately without expecting it to repeat, or hold a broad expectation without retrieving a particular episode. Waters and colleagues examined both coherent discourse about childhood care and secure base script knowledge in a longitudinal sample. The two indicators made unique associations with different later relationship outcomes. This supports distinguishing representational measures, while warning against saying that one interview or narrative fully captures a person's relational knowledge.
Continuity and change
Waters and colleagues followed a small group from infant attachment assessment to an adult attachment interview about twenty years later. They reported substantial, not complete, agreement in secure versus insecure classification, and changes were associated with major negative life events. This is evidence of continuity under some conditions and change under others. The study's sample was small and socially narrow. It should not become a numerical destiny claim about all children or a prediction for any individual infant.
A later six-year study by Waters, Yang and colleagues followed children into adolescence using both questionnaire and narrative-based measures. Its patterns favored a prototype model of attachment stability in that sample: a relatively persistent component under temporary variation. But a prototype is not a prison. The authors examined moderators, and the sample was a defined White Western European group. The finding supports some durable organization of expectations while leaving room for relationship experience and developmental change.
Ammaniti and colleagues assessed attachment representations in a small Italian sample at ages ten and fourteen. Security showed considerable stability across that transition, with variation among categories. Small samples and interview classifications limit the generality of exact percentages. The study is useful because it examines a developmental interval when peers, autonomy and family roles are changing. The model may persist while the situations that activate it change, and a young person's relationship to caregivers can be reorganized rather than simply replayed.
Change is theoretically expected when important relationships provide new information. A reliable partner, friend, therapist or caregiver may challenge an expectation that help will fail. A threatening relationship may produce the opposite. The evidence does not specify a guaranteed dose of new experience or a universal direction. Institute synthesis: the practical question is what repeated, credible experiences are available and whether the person can use them safely. Telling someone to 'rewrite their model' without changing relational conditions makes the concept sound more controllable than it is.
General and relationship-specific models
Fraley and colleagues developed a questionnaire that measures attachment orientations across distinct relationship contexts. In their studies, relationship-specific scores often predicted corresponding interpersonal outcomes better than broad scores, while broader scores related more strongly to general personality features. This pattern challenges the assumption that every adult has one uniform template applied identically to all people. It also shows why an average attachment score can obscure meaningful variation between a parent, partner, friend and mentor.
Specificity is not the same as inconsistency or pathology. A person may trust a longstanding friend and remain cautious with a new supervisor for good reason. A relationship-specific model can reflect evidence about that relationship. The research question is whether the representation is calibrated to current experience and how it influences behavior. A broad tendency can still shape what a person notices across contexts. The two levels can coexist. Researchers should state which level their measure captures before making a claim about 'attachment security.'
The distinction matters for intervention. If a person feels safe with one relationship and not another, global advice to become more trusting is too coarse. It may be better to examine the history and present reliability of the specific relationship. In a harmful situation, distrust can be protective rather than a distorted working model. Institute synthesis: an ethical use of attachment language starts with the quality of care actually provided. It does not ask someone to override evidence of unreliability merely to fit a preferred classification.
How models are measured
The Adult Attachment Interview evaluates how a person discusses attachment-relevant experiences, including coherence of discourse. A secure base script assessment asks for narratives from which researchers infer organized knowledge of effective support. Self-report questionnaires ask about anxiety and avoidance in particular relationships. These methods are related but not interchangeable. Each has a task, coding rules and limits. A clinical impression drawn from casual conversation is not equivalent to a validated assessment, and even a validated score is not an exhaustive portrait.
Measurement also raises cultural and linguistic questions. Narrative coherence may depend on conversational conventions; a questionnaire's phrase 'depend on others' may carry different meanings across communities. Cross-cultural use requires evidence that a measure retains its intended structure and interpretation. The longitudinal studies cited here come from specific samples and cannot establish that one form of expression is universal. A secure base can be enacted through different patterns of shared care, proximity and communication. Researchers should examine function before imposing one outward style.
Models are often described in terms of self and other. A person might expect others to respond and feel comfortable seeking help, or anticipate rejection and avoid asking. Such descriptions are hypotheses about patterns, not labels to assign from a single text message or relationship conflict. Anxiety and avoidance questionnaires may reveal dimensions, but they do not identify the cause of a present disagreement. A model can guide attention while the other person's behavior also matters. A good explanation examines both sides of the interaction.
What the evidence supports
Longitudinal findings support the idea that some attachment-related representations carry information across time. The Steele cohort connects observed early caregiving with later script knowledge; the Waters studies show continuity and measure-specific relations with later relationships. The Fraley work shows that specificity matters. None isolates one causal pathway conclusively, because family circumstances, temperament, later events and measurement choices also contribute. The evidence is stronger for probabilistic patterns than for deterministic narratives about one childhood cause.
Adult relationship outcomes are likewise varied. Waters and colleagues found that coherent discourse and secure base script knowledge had different unique associations with observed romantic and parent-child measures. That specificity is scientifically useful. It discourages a sweeping claim that one representation predicts all intimacy, parenting and mental health outcomes. A relationship can be good in one dimension and difficult in another. Researchers should specify whether the outcome is self-reported satisfaction, observed behavior or a child's attachment classification.
An internal working model should not be confused with an explicit belief that is always available for verbal report. People can sometimes describe an expectation, but they may enact a pattern before naming it. Conversely, a stated belief may not match behavior under stress. Multiple methods can test whether the proposed representation helps explain action beyond what is already known from the current relationship. If it adds no predictive value, the explanation may be too broad. The construct remains useful when it produces discriminating, testable predictions.
Developmental timing also deserves restraint. The model is not completed at a single birthday. Children acquire language, encounter new caregivers and peers, and gain new ways to interpret past events. Adolescents negotiate increasing independence while still needing support. Adults can encounter relationships that confirm or challenge existing expectations. The cited longitudinal studies measure selected points along that process, not every intervening experience. A claim of stability over years therefore means that a statistical pattern persists amid change, not that the person’s interpretation never evolves. Researchers should describe the interval and measure before calling a model stable.
Applications with restraint
For a caregiver, the concept highlights the importance of reliable, responsive support over repeated interactions. It should not be turned into pressure to respond perfectly at every moment. A relationship includes repair after misunderstanding, different caregivers and changing developmental needs. Institute synthesis: the aim is a pattern in which help is accessible and exploration remains possible, not a flawless performance. Services that support caregivers should attend to time, safety and resources. A caring intention cannot substitute for conditions that permit care.
In adult relationships, it can be useful to ask what each person expects when distressed and whether those expectations fit the current interaction. This can improve description, but it is not a diagnosis or a license to blame one partner for all conflict. A person's hesitation may reflect a present breach of trust; a partner's withdrawal may reflect exhaustion rather than a stable attachment model. Professionals need assessment and context before using an attachment formulation. A public Learning Hub entry can explain the research without prescribing relationship repair.
What the evidence does not show
The evidence does not show that early attachment fixes destiny, that every later choice reveals one hidden model, or that a narrative measure directly reads an internal structure. Longitudinal associations can coexist with change; classification stability varies by sample, interval and method. Early experience is important without being the only influence. The cited studies also cannot establish a universal cultural script for how support must look. Claims about a particular person require careful attention to their relationships and circumstances.
Nor does the evidence show that distrust is always a mistaken residue of childhood. A current relationship may be unreliable or unsafe. Applying the model responsibly means testing the fit between expectation and present evidence, not instructing someone to trust against their judgment. In some cases, changing the environment or leaving a harmful relationship is more relevant than reframing an expectation. This is an Institute synthesis consistent with the construct's emphasis on experience, not a clinical treatment claim.
Internal working models are most useful as a bridge between episodes and patterns. They ask how repeated care becomes a working expectation, how that expectation guides later interpretation, and how new experience may revise it. The evidence supports continuity with meaningful limits, multiple levels of specificity and more than one way to measure representation. Keeping those distinctions visible makes the concept informative without turning it into a verdict on a person's past or future.
Continue exploring
Research references
- Waters, E., Merrick, S., Treboux, D., Crowell, J., & Albersheim, L. (2000). Attachment security in infancy and early adulthood: A twenty-year longitudinal study. Child Development, 71(3), 684–689. View source
- Ammaniti, M., van IJzendoorn, M. H., Speranza, A. M., & Tambelli, R. (2000). Internal working models of attachment during late childhood and early adolescence: An exploration of stability and change. Attachment & Human Development, 2(3), 328–346. View source
- Fraley, R. C., Heffernan, M. E., Vicary, A. M., & Brumbaugh, C. C. (2011). The Experiences in Close Relationships-Relationship Structures questionnaire: A method for assessing attachment orientations across relationships. Psychological Assessment, 23(3), 615–625. View source
- Steele, R. D., et al. (2014). Caregiving antecedents of secure base script knowledge: A comparative analysis of young adult attachment representations. Developmental Psychology, 50(11), 2526–2538. View source
- Waters, T. E. A., et al. (2018). Adult attachment representations and the quality of romantic and parent-child relationships: An examination of the contributions of coherence of discourse and secure base script knowledge. Developmental Psychology, 54(12), 2371–2381. View source
- Waters, T. E. A., Yang, R., Finet, C., Verhees, M. W. F. T., & Bosmans, G. (2022). An empirical test of prototype and revisionist models of attachment stability and change from middle childhood to adolescence: A 6-year longitudinal study. Child Development, 93(1), 225–236. View source