RESEARCH DOSSIER
Relational Function Redistribution
Abstract. Relational Function Redistribution investigates the changing allocation of psychological and communicative work among human partners, artificial systems, and institutions. This structured critical narrative monograph defines the proposed construct through function-specific events and longitudinal configurations; reconstructs its foundations in social support, attachment, transactive memory, distributed cognition, cognitive offloading, agency, human–computer social response, and trust in automation; and critically reviews experimental and observational evidence on AI companionship, loneliness, disclosure, clinical conversational support, and platform disruption through October 2026. The synthesis distinguishes redistribution of functions from overall social substitution, identifies competing causal explanations and measurement errors, and proposes repeatable allocation matrices, dyadic longitudinal designs, controlled interventions, and cross-platform replication. Existing studies support component mechanisms while leaving the unified construct and its population-level causal consequences open to independent empirical validation. This work presents a scholarly critical synthesis and research agenda rather than a systematic review, meta-analysis, or original experiment.
Keywords: relational function redistribution; postsubjective psychology; human–AI relationships; relational support; social support; emotional reassurance; social interpretation; transactive memory; cognitive offloading; autobiographical memory; conflict rehearsal; distributed cognition; AI companionship; relational continuity; social substitution; complementarity; longitudinal psychology; construct validity; Artificial Era
Introduction: The Distribution of Relational Functions
Relational Function Redistribution names the changing allocation of psychological and communicative work across people, artificial systems, and institutions over time. A relationship is sustained through activities: providing reassurance, interpreting uncertain events, remembering shared histories, preparing difficult conversations, coordinating plans, and regulating expectations. Some activities are performed by one person, others jointly, and still others through records, organizations, or technical services. Artificial conversational systems add participants and interfaces through which this work can be requested, rehearsed, stored, and returned. The scientific object is the allocation and movement of these functions, observed through behavior and reports across repeated situations. The term here specifies a research construct within the Postsubjective Psychology program, subject to independent operational validation rather than assumed as a previously established diagnostic category.
Function redistribution poses a longitudinal question. A person may initially consult a friend for emotional interpretation, later practice with an artificial assistant before speaking to that friend, and eventually rely on a stored conversation for autobiographical recall. The contact count with the friend might remain stable while the functions of the relationship have changed substantially. Conversely, an artificial resource may facilitate a new conversation with another person, extending rather than displacing social interaction. A research design therefore needs to record which function was performed, by whom, with what degree of involvement, for what purpose, and with which later consequences. The distribution is a time-varying relational arrangement, not a fixed preference for one partner type.
This monograph integrates research on social support, attachment, distributed cognition, transactive memory, cognitive offloading, agency, anthropomorphism, disclosure, automation reliance, and recent studies of AI companionship. It treats the contributions of those literatures as converging but distinguishable evidence. The method is a structured critical narrative synthesis, grounded in original studies and major theoretical works. No exhaustive search protocol, pooled effect estimate, experimental dataset, or claim of validated causal effect for the named construct is presented. Its contribution is to specify a measurable scientific problem and evaluate the strength of existing evidence for each proposed component. Cohen & Wills (1985) Hutchins (1995) Risko & Gilbert (2016) Zhang et al. (2026)
Research Object and Scientific Boundaries
The unit of analysis is a function-allocation event embedded in a sequence. An event links an initiating need, a task or psychological purpose, a provider or tool, an interaction process, and an observable consequence. For example, after an interpersonal disagreement, a participant may ask an AI system to reconstruct the other person's likely perspective. That event involves interpretation; a later call with the person may involve reconciliation; a subsequent diary entry may document whether the system's interpretation was used, rejected, or revised. The same actor can perform multiple functions, and one function can be distributed across several actors. Function-level measurement allows the investigator to examine these combinations without assuming that all use constitutes attachment, support, or dependency.
Relational functions should be operationalized as distinguishable activities and outcomes. Reassurance refers to perceived or enacted reduction of threat and uncertainty in a relational context. Interpretation refers to assigning meaning to another person's conduct or intentions. Remembering refers to retrieval, organization, or preservation of shared personal information. Conflict rehearsal refers to preparing a communicative act before an interpersonal exchange. Planning refers to coordinating commitments, decisions, and task execution. Support includes emotional, informational, appraisal, and instrumental provisions, categories that may overlap in practice. Each definition requires corresponding observations, participant reports, or task measures, with evidence that participants interpret the category similarly across human and technical settings.
This object has a precise boundary with Social Substitution and Complementarity in AI Relationships. Substitution research asks whether human connections, interactions, or support are displaced or augmented. Redistribution research tracks the migration, duplication, specialization, and recombination of particular relational functions, including cases with unchanged network size. It also differs from attachment research, which examines the organization and regulation of bonds, and from cognitive offloading, which studies reliance on external resources for cognitive tasks. Redistribution integrates function-specific outcomes across these literatures; it does not replace their constructs. Psychological consequences, including loneliness, confidence, coordination, intimacy, and conflict, are dependent variables requiring separate assessment. Feeney & Collins (2015) Wegner (1987) Risko & Gilbert (2016)
From Distributed Cognition to Human–Artificial Configurations
Hutchins examined cognition as an organized process extending across people, representations, tools, and institutional routines. Navigational activity supplied a concrete example of how representational transformations and coordinated actions can accomplish tasks whose explanation exceeds any isolated individual's internal computation. Hollan, Hutchins, and Kirsh subsequently articulated distributed cognition as a foundation for human–computer interaction research, including the propagation of representations through artifacts and time. Relational function redistribution adopts the methodological lesson that investigators should trace transformations across the actual work system. It does not claim that every component in that system has the same psychological capacities. Hutchins (1995) Hollan et al. (2000)
Clark and Chalmers argued philosophically that under certain conditions external resources can play roles ordinarily attributed to internal cognitive processes. The extended mind debate opened questions about availability, reliable access, trust, and coupling between person and artifact. In relational research, a personal conversational history, shared calendar, private journal, or AI conversation can become an external source in the ongoing organization of remembering and deciding. A researcher must still establish when and how the resource changes performance and experience. The philosophical possibility of an extended process does not establish that a particular AI service actually improves memory, improves relationships, or retains continuity across updates. Clark & Chalmers (1998)
The relevant configuration can contain a human participant, one or more human partners, an artificial interface, the provider's platform policies, stored records, and institutional norms. A configuration changes when a conversation is saved, a model is updated, an account is suspended, another person is given access, or a once-private exchange enters a family decision. These transitions can reorganize the function of an apparently stable tool. Useful records therefore describe system version, memory state, access permissions, interaction modality, and the participant's awareness of each feature. These are potential explanatory variables, not decorative technical metadata. They make it possible to study how a relational function becomes easier, harder, or qualitatively different to perform.
Transactive Memory and the Coordination of Knowing
Wegner's formulation of transactive memory analyzes how people in close groups organize knowledge through differentiated expertise and information about who knows what. Collective functioning can depend on locating a source as much as on retaining the information personally. A household may distribute appointments, shared history, technical expertise, and emotional memories among members and documents. As computational assistance enters this arrangement, the empirical question is whether people change where they seek information, which source they regard as authoritative, and how they verify its response. A conversational system may become an index into personal records, an external summarizer, or a source of erroneous but plausible reconstructions. Wegner (1987)
Research on digital memory gives an experimentally grounded analogy. Sparrow, Liu, and Wegner examined how expectations of later access to information altered subsequent remembering and memory for where material could be found. Risko and Gilbert's review situated this phenomenon within the wider study of cognitive offloading, including choices to store, externalize, or delegate cognitive operations. These findings support the plausibility of function migration between internal remembering and available resources. Their experiments do not directly measure a person's intimacy, autobiographical dependence on AI companionship, or downstream changes in family communication. Extension into those domains requires purpose-built evidence. Sparrow et al. (2011) Risko & Gilbert (2016)
Relational remembering has a special content problem: a personal event is often contested, emotionally salient, and jointly interpreted. An artificial summary can preserve one narrative while omitting contextual qualifiers, thereby altering the version that participants subsequently discuss. Researchers should compare source transcripts with later recalled accounts, evaluate omission and confabulation rates, and examine whether memory use strengthens or weakens interpersonal verification. The external archive's persistence may support continuity while creating privacy and access vulnerabilities. A robust study distinguishes retrieval accuracy from confidence, perceived autobiographical continuity, and relational consequences. These outcomes can move in different directions, making a single global measure of 'memory help' scientifically insufficient.
Reassurance and Emotion Regulation
Reassurance is a relational function when a person seeks help assessing threat, uncertainty, rejection, or anticipated social consequences. It can be delivered through listening, affirmation, practical feedback, shared action, and reminders of a person's competencies. In close relationships, reassurance carries information about the partner's availability and investment, whereas an AI response is shaped by system design and current conversational inputs. These differences can affect what a recipient believes the response guarantees. The first measurement task is to identify the sought function: immediate emotional soothing, confidence in a plan, validation of experience, or advice about contacting another person. A questionnaire that counts supportive messages without specifying these ends confounds distinct psychological processes.
Attachment research offers a historically important framework for understanding reassurance within human bonds. Hazan and Shaver applied attachment concepts to romantic relationships, connecting expectations of safety and proximity with patterns of adult closeness. Within their domain, reassurance derives meaning from the history of interaction and the partner's actual capacity to respond to future needs. Artificial interfaces may become salient sources of predictable conversation and perceived responsiveness, but equivalent self-reported comfort does not establish equivalent attachment mechanisms or obligations. The testable claim is narrower: access to a reliably responsive interface may alter when, how often, and from whom reassurance is sought. Hazan & Shaver (1987)
Redistribution may unfold in several sequences. A person might consult an assistant to reduce distress, then return to a human conversation with more clarity; a different person might use the assistant after withdrawing from human support; another might compare the assistant's reassurance against a partner's interpretation. These patterns have different implications for autonomy and relationship maintenance. Event-contingent diaries can identify the precipitating episode, self-rated distress, provider choice, immediate change, subsequent contact, and retrospective evaluation. A causal study would experimentally vary access or response features while monitoring genuine interpersonal outcomes, since a decrease in momentary distress alone cannot reveal the full effect on relational organization. Cohen & Wills (1985) Feeney & Collins (2015)
Interpretation, Perspective Taking, and Sensemaking
Interpretation is the construction and revision of meaning in response to an ambiguous social episode. In relationships, people infer intentions from language, behavior, timing, and shared history; they also test those interpretations by seeking clarification. An artificial assistant can propose alternative explanations for a partner's behavior or summarize competing perspectives. In the resulting configuration, an interpretive task may be performed sequentially by the person, the assistant, and the partner. The research object is the transfer and recombination of inferential labor. Observables include the initial interpretation, alternative explanations generated, confidence adjustments, information later verified, and whether the individual seeks direct communication.
A useful theoretical anchor is Bandura's account of human agency, in which people act through individual, proxy, and collective pathways. Requests for external interpretation can function as assistance in one's own decision process, while stronger reliance can approach proxy decision making when the recipient permits a source to set judgments on their behalf. This distinction becomes testable through decision ownership, acceptance of advice, willingness to inspect evidence, and readiness to override an erroneous interpretation. A respondent who says an assistant 'understood the situation' may be reporting relief or fluent wording rather than an objectively accurate model of the absent partner's motives. Research must measure these separately. Bandura (2001)
Interpretive redistribution involves hazards that grow with asymmetrical information. An assistant often receives one participant's version of an interpersonal conflict; the absent person's intentions and circumstances may remain inaccessible. The system's response can therefore strengthen one narrative, introduce unsupported inferences, or encourage useful perspective taking, depending on design and use. Paired-partner studies should compare privately reported accounts, expose participants to controlled advice, and later measure whether clarification with the actual partner occurred. Such designs can evaluate interpretive calibration without requiring researchers to declare one participant's feelings invalid. The proposed mechanism is a change in the source, timing, and authority of relational explanations, requiring direct observation before general causal conclusions can follow.
Autobiographical Memory and Relational Continuity
Autobiographical remembering maintains narratives of a person's life and the events through which relationships develop. Partners frequently remember different details, preserve photographs and messages, and refer to shared records to settle dates or reconstruct experiences. Artificial conversational systems make it possible to search and summarize such records, but the reliability of a memory function depends on which records exist and whether retrieval distinguishes evidence from generated inference. In a longitudinal study, the researcher can observe a movement from recollection by a person to recollection aided by a stored transcript or generated summary. The changed location of information matters alongside the accuracy and emotional consequences of retrieval. Wegner (1987) Sparrow et al. (2011)
An external conversational memory can change expectations about continuity. A user who expects a system to recall personal preferences may disclose less background in future sessions and interpret remembered details as a sign of stability. When the system fails to retrieve a detail, that failure may be experienced as a loss of practical support or relationship continuity. Distinguishing these experiences requires combining behavioral memory tests with self-report measures of confidence, trust, perceived familiarity, and continuity. The researcher should record actual technical capabilities rather than rely on a vendor's generic promise of 'memory.' The same feature label may refer to context-window retention, selective saved facts, retrieval from documents, or personalized model settings.
Autobiographical function redistribution intersects the governance of private information. Records created in intimate circumstances may later be used to answer unrelated questions, train services, or influence recommendations, depending on contractual arrangements and technical controls. A valid design documents consent, data provenance, retention, access rights, deletion procedures, and third-party interests. Accuracy measures should compare generated descriptions with original records and participant-confirmed events, while separate measures assess willingness to rely on the generated reconstruction. Such work can establish whether conversational memory reduces ordinary effort, changes source monitoring, or alters the social practice of remembering together. These mechanisms remain empirical hypotheses rather than automatic consequences of using a memory-capable system. Risko & Gilbert (2016)
Conflict Rehearsal and Communicative Preparation
Conflict rehearsal consists of practicing an anticipated communication before the encounter takes place. People may rehearse with friends, therapists, diaries, or imagined interlocutors; conversational AI offers another interactive environment for exploring wording, timing, and likely reactions. The relational function is preparation for interaction, so relevant outcomes include the subsequent conversation, not simply satisfaction with the rehearsal. A useful protocol distinguishes drafting a message, role-playing alternative responses, emotionally processing a grievance, and deciding whether direct contact is appropriate. These are different tasks with distinct behavioral endpoints. A study might preserve anonymized drafts, independently code their interpersonal tone, and compare them with the final communication when participants consent.
An AI system may make rehearsal more accessible through on-demand availability, rapid revision, and low perceived social cost. Experimental work on self-disclosure to virtual humans suggests that expectations about a conversational partner can influence what people are willing to reveal. Lucas and colleagues found conditions in which virtual-human interactions increased willingness to disclose, although the study did not evaluate enduring human–AI bonds or actual conflict resolution in close relationships. This evidence supports testing perceived evaluation threat as a mediator of rehearsal choices. The transfer from virtual disclosure to successful interpersonal communication remains a new question requiring observed behavior and follow-up. Lucas et al. (2014)
Potential mechanisms include reduced anticipatory anxiety, improved formulation of requests, increased perspective taking, and, conversely, repeated rehearsal that sustains rumination or rehearses a narrow narrative. These possibilities demand differentiated measurement. A randomized comparison could assign participants to structured self-reflection, human-mediated rehearsal, or AI-assisted rehearsal while holding practice time constant. Blind observers could assess clarity and responsiveness in subsequent conversations, while participants and partners would report emotional outcomes. Without partner data, a perceived improvement in communicative confidence cannot establish that the interpersonal event improved. The central scientific issue is when an artificial rehearsal environment helps transfer skills to a real relationship and when it changes only the subjective readiness to engage.
Planning, Delegation, and Proxy Agency
Relational planning includes the coordination of schedules, household commitments, caregiving, conflict-repair steps, and shared decisions. Coordination frequently uses calendars, lists, institutional staff, and task-specific expertise. An AI assistant can generate options, draft plans, track preferences where permitted, and help a person sequence a complicated discussion. Function redistribution occurs when the source of planning work changes or when an existing allocation becomes more fine-grained. For example, an assistant may handle appointment reminders while partners retain value judgments and final decisions. The distinction between preparing options and exercising authority is central to measuring whether technical mediation supports or weakens shared agency.
Bandura's social cognitive treatment of agency differentiates direct personal action, relying on others to act on one's behalf, and collective agency achieved through coordinated interdependence. These distinctions help specify the roles of artificial systems. A system can be used as a planning instrument, delegated a bounded task, or regarded as an apparently authoritative decision adviser. The researcher can examine who initiates the action, who may revise it, who is responsible for outcomes, and whether affected people gave informed permission. Proxy assistance becomes particularly significant when a conversation addresses a third person's privacy or when an automated recommendation enters a decision with consequential effects. Bandura (2001)
Appropriate reliance depends on accuracy, transparency, and the user's capacity to detect errors. Lee and See's framework for trust in automation identifies trust as relevant to whether users employ a system under uncertainty. In relational contexts, confidence may be shaped not only by objective performance but also by perceived warmth, consistency, and familiarity. Researchers should experimentally vary demonstrable accuracy and socially persuasive cues separately, measuring calibration and the consequences of misplaced reliance. A successful planning function is one in which technical support improves coordination while decisions and responsibilities remain traceable. The broader research question is how allocation changes when systems begin to act on external tools and shared information rather than only supply conversational suggestions. Lee & See (2004)
Anthropomorphism and Social Response
Anthropomorphism is a process through which people attribute humanlike properties, intentions, or experiences to a nonhuman target. Epley, Waytz, and Cacioppo developed a three-factor account that relates anthropomorphism to accessible knowledge about humans, motives to explain or control an agent, and motivations for social connection. These factors offer hypotheses about variation between users rather than a single inevitable response to artificial conversation. In redistribution research, humanlike presentation may influence which relational functions people entrust to a system and how strongly they weight its responses. A fluent conversational style may promote interpretation and reassurance use even when actual expertise and interpersonal accountability are unchanged. Epley et al. (2007)
Nass and Moon documented socially patterned responses to computers, showing that characteristics of an interface can elicit conventions ordinarily observed in human interaction. The findings are historically important because they demonstrate that social responses can arise in human–computer settings without requiring deliberate belief that a computer is literally a person. Their paradigms establish a laboratory foundation for studying interface framing and user behavior, while their dates and technical contexts limit direct inference about contemporary personalized AI relationships. Modern studies should isolate the effects of system accuracy, conversational responsiveness, persistent memory, anthropomorphic labels, and user expectations, rather than treating them as one undifferentiated 'humanlike' factor. Nass & Moon (2000)
Attribution has practical consequences for function allocation. A person who perceives an assistant as patient, confidential, or unusually understanding may assign the assistant a larger share of emotionally sensitive interpretation than a less anthropomorphic tool. That allocation can be measured with choice tasks that offer otherwise identical providers under different framing conditions, followed by real-world use diaries. The investigator must distinguish a short-term change in provider choice from a sustained shift in relationship structure. It is also useful to assess participants' explicit beliefs about what the system can remember, whether it knows an absent third party, and who bears responsibility for advice. These beliefs may mediate the relation between presentation and use.
Disclosure, Perceived Responsiveness, and Trust
Self-disclosure is a communicative practice through which personal information becomes available to an interaction partner. Disclosure may support understanding and closeness in human relationships, but its consequences depend on perceived safety, reciprocity, response quality, and control over the information. Artificial interfaces can lower some social barriers to initial disclosure by avoiding face-to-face evaluation, yet the recipient's assumptions about privacy and storage may be inaccurate. Lucas and colleagues' virtual-human experiment indicates that an interaction frame can alter disclosure willingness. It does not demonstrate that an automated system can maintain a reciprocal human bond or that disclosure is invariably beneficial. The redistribution question is whether disclosure-related functions move between confidants, journals, professionals, and AI interfaces, and under which conditions. Lucas et al. (2014)
Trust in an automated partner includes beliefs about its competence and the reliability of its outputs. Lee and See emphasize the design of appropriate reliance: a user should employ automation according to its actual capabilities in the task at hand. In relational advice, perceived responsiveness may be high when a system acknowledges an emotion and offers coherent suggestions, even when it lacks independently verified information about the people involved. Trust calibration can be measured by presenting advice of known reliability and observing acceptance, verification, and revision. The outcome of interest extends from first impressions to repeated choices across varying stakes and accuracy histories. Lee & See (2004)
Disclosure and trust can reinforce each other through feedback. A user who receives a satisfying response may disclose more and return more frequently, changing the amount of personal information available to the system. The system can then produce more contextually specific language, which may strengthen the user's confidence, including when the additional detail does not improve factual accuracy. A longitudinal design should record disclosure depth, voluntary use, system memory capability, objective correctness where assessable, and participant expectations. Qualitative interviews can illuminate why particular interactions were experienced as safe, while independent coding and event logs can assess what actually changed. The causal interpretation remains conditional on distinguishing self-selection from the effects of repeated access.
AI Companionship and Social Network Ecology
AI companionship encompasses uses in which an artificial interface is experienced as a source of conversation, support, familiarity, or perceived closeness. Such uses occur within a wider social ecology of friends, partners, family members, colleagues, professional relationships, and institutions. Research on function redistribution examines the relation between these sources rather than assigning a single global effect to the AI system. A participant may gain access to informational support while maintaining human emotional support; another may shift personal disclosure away from a close friend while continuing ordinary contact; a third may report companionship with little change in measured social network behavior. Each pattern entails a different allocation of functions.
Zhang and colleagues' 2026 study combined survey data from U.S. adult users of Character.AI with donated conversational material from a smaller subsample. The work links patterns of companion use and psychological well-being in a naturalistic population, permitting study of associations, conversational characteristics, and heterogeneity. It cannot establish that the platform itself caused a given user's lower well-being, because people with different initial needs may select different patterns of use. This distinction is important: observed emotional intimacy with a system may be a response to preexisting isolation, a contributor to later experience, or both. The scientific account must preserve these possibilities until appropriate designs can distinguish them. Zhang et al. (2026)
Human network outcomes should be disaggregated. The quantity of social contact, perceived connection, perceived emotional isolation, availability of instrumental help, tie strength, relationship satisfaction, and conflict management describe overlapping but independent dimensions. A function-allocation diary can connect those outcomes to uses of artificial conversation for particular purposes. Social network methods can track whether support roles are concentrated in one partner or distributed across multiple sources. The research should also identify differences between personal companions, task assistants, professional services, and creative role-play systems. A classification built only around the interface type risks overlooking the activity for which it is chosen.
Substitution, Complementarity, and Function Migration
Substitution refers to an instance in which use of one source replaces use of another source for a specified function. Complementarity refers to a case in which the presence of an additional source increases, enables, or improves the use of a different source. A mixed arrangement can contain substitution in one function and complementarity in another. For example, a person may use AI for first-pass message drafting while increasing direct communication with a partner. A global count of human interactions would capture the latter change but would miss the transferred drafting function. A multidimensional model should therefore describe simultaneous roles rather than classify the entire user as either substituting or augmenting relationships.
Function migration adds temporal direction to the analysis. The main observations are an earlier allocation, an intervening change in access or circumstances, and a later allocation, with uncertainty concerning why the change occurred. Migration may involve full transfer, temporary delegation, alternating use, joint performance, or recovery of a previously externalized function. A function-allocation matrix can record the proportion of episodes handled by each source, provided the coding scheme allows multiple actors per event and preserves absolute frequency. Without absolute frequencies, a change in shares could reflect the disappearance of a task rather than a genuine transfer of responsibility. Researchers must therefore measure task demand as well as provider choice.
Causal interpretation requires counterfactual comparisons. A person who increasingly turns to AI might be coping with a disrupted friendship, an exhausting work schedule, or a newly available platform. Exposure is time varying and influenced by prior outcomes, so simple before-and-after correlations are vulnerable to confounding. Experiments that vary system access, quality, or memory may help identify short-term reallocations, while longer studies with partner reports are necessary to evaluate relationship consequences. Folk and Dunn's 2026 study provides an example of the bidirectional problem: baseline social circumstances and later chatbot use may be related in ways a one-direction account cannot adequately describe. Folk & Dunn (2026)
Platform Change, Rupture, and Relational Dependence
Artificial relational functions are mediated by platforms whose features and continuity can change through updates, moderation policy, model replacement, service outages, and account restrictions. A user's arrangement of support or remembering may become dependent on that institutional infrastructure. A sudden change can affect access to prior records, the wording and apparent familiarity of responses, or whether intimate conversation remains possible. These events are opportunities for scientific study because the researcher can observe consequences of changes to an existing configuration. They also require careful documentation: a general notice of an update cannot substitute for a technical record of which features actually changed and when.
De Freitas and colleagues' 2026 research on mourning the loss of AI companions examined reports following disruptions and combined large-scale online material with survey evidence. The study documents that some users experience meaningful distress when AI companionship is interrupted or altered. Such reactions are psychologically consequential reports of experienced loss. Inference about magnitude in the general population must consider self-selection, the observational character of online discourse, and differences between people who post publicly and those who do not. The findings support the importance of studying continuity and rupture, while leaving the frequency, mechanisms, and causal determinants of these responses open to further research. De Freitas et al. (2026)
A function-specific account clarifies what rupture removes. A participant may lose access to an emotionally reassuring conversational style, an indexed personal history, a predictable rehearsal routine, or an individualized planning arrangement. These losses can co-occur, but they should be recorded separately because their consequences and possible substitutes differ. Natural experiments around platform changes can collect data before and after the event, document alternative support access, and examine whether participants reconstruct functions with other people or systems. Institutional accountability enters the analysis when the system's provider determines data portability and changes to a resource on which users have learned to depend.
Experimental Evidence on Support and Loneliness
An experimentally controlled evaluation can distinguish immediate perceived support from changes in loneliness and real social behavior. Different outcomes require different durations and comparators. Fitzpatrick, Darcy, and Vierhile tested an automated cognitive behavioral conversational agent with young adults reporting symptoms of depression and anxiety in a randomized trial, providing an early example of a defined intervention evaluated against a comparator. Its short duration, small sample, and intervention context constrain the implications for everyday companionship or long-term relational functions. One investigator's disclosed financial interest in the company developing the intervention is relevant to independent appraisal of this trial. The trial supports the feasibility of studying automated conversational support with experimental methods; it cannot establish broad effects on attachment, social substitution, or the allocation of support among close partners. Fitzpatrick et al. (2017)
A 2026 preregistered study by Li and colleagues compared interaction with a randomly assigned human peer, a supportive chatbot, and a journaling condition among university students. The study directly addresses whether distinct support providers have similar consequences for loneliness over a short period. Its design strengthens causal inference relative to a cross-sectional companion-user survey because conditions were assigned rather than merely observed. The authors reported stronger reductions in loneliness in the peer-interaction condition than in the chatbot condition. That result concerns the studied population, protocols, duration, and measures, and does not mean all AI assistance is ineffective for other functions. It illustrates why the relational work being requested must be distinguished from generalized availability of interaction. Li et al. (2026)
A longitudinal controlled chatbot study reported by Fang and colleagues in 2025 examined different interaction conditions and psychosocial outcomes over four weeks. The investigators did not observe a simple, uniformly beneficial or harmful pattern attributable to randomized conversational conditions. They also analyzed associations between greater voluntary use and psychosocial measures, findings whose interpretation depends on reverse causation and selective engagement. The publicly available work is a preprint, so its claims should retain that status until a verified peer-reviewed version is established. Across these studies, experimental evidence supports testing specific interventions, conditions, and populations. It does not yet identify a universal function-redistribution trajectory for all users. Fang et al. (2025)
Longitudinal Evidence and Bidirectional Selection
Repeated observations help identify whether function allocation precedes or follows changes in social life. Folk and Dunn examined associations between AI companionship and loneliness across time, including the possibility that initial social experience predicts later chatbot use. Their findings separated emotional isolation from broader social connection and showed why those outcomes should not be collapsed into a single metric. A participant may feel emotionally isolated while maintaining a network of contacts, or report broad social connectedness with little confidence in available reassurance. Such distinctions are essential for specifying when a relational function has moved and whether that movement is associated with improvement or deterioration in a particular domain. Folk & Dunn (2026)
Observational companion data collected by Zhang and colleagues provide complementary evidence about naturally occurring use, disclosure, and well-being. Naturalistic records reveal behavior outside artificial laboratory tasks but are exposed to self-selection, platform-specific norms, and uneven data donation. A participant who selects companionship-oriented use may differ from a participant seeking technical assistance before any exposure is measured. Within-person repeated measures can reduce some stable individual confounding, yet changes in life circumstances, relationships, or platform functions remain alternative explanations. The strongest observational programs would follow both function-allocation events and relevant social outcomes before as well as after major transitions. Zhang et al. (2026)
Longitudinal designs also face attrition and reactivity. People experiencing distress or service interruption may be more likely to stop recording events, and asking users to categorize every conversation could change how they use their supports. Passive logs can reduce recall burden but rarely reveal whether a message was experienced as reassurance, planning, or mere entertainment. Combining logs with short event-contingent questions, periodic network inventories, and consenting partner reports improves interpretability. Analysis plans should distinguish within-person from between-person associations and specify time scales: same-day changes, weekly adaptation, and sustained allocation across months may arise through different processes. Robust inference depends on measuring the initial situation before classifying subsequent shifts as effects of AI use.
Clinical Support, Care Boundaries, and Accessibility
Clinical and counseling contexts provide important but bounded evidence about mediated support. A structured digital mental-health intervention may teach coping skills, facilitate self-monitoring, or provide conversational exercises. Its effectiveness must be assessed using relevant clinical outcomes, comparators, and safety monitoring. The Woebot randomized study was a preliminary intervention test with a specific young-adult population, not a validation of an unrestricted social companion as a therapist. Generalizing from such trials to clinical care requires evidence about diagnostic populations, duration, adverse outcomes, human oversight, and the regulatory status of the intervention. The research value for redistribution is the possibility that task-specific support can be allocated to a tool while human professionals retain other functions. Fitzpatrick et al. (2017)
Accessibility can influence redistribution by changing the relative cost of obtaining a function. A person who cannot contact a human supporter immediately may use an always-available conversational system to rehearse a request or organize information. Another may need assistive technology to make communication possible at all. These cases should be evaluated against the relevant accessible alternatives, including peer services, written tools, and professional support. A system's high availability may help with timely preparation even when it cannot offer the contextual knowledge, accountability, or practical intervention supplied by people. Accessibility claims require direct evidence among the affected users, not general deductions from the existence of a chat interface.
Clinical interpretation is particularly sensitive to emotional dependence, distress after interruptions, and misleading advice. Studies should record the system's stated scope, actual safety behavior, routing to professional or emergency support where appropriate, and whether reliance delays effective care. Outcomes may include symptom change, functional impairment, service engagement, adverse events, and quality of life, each requiring validated instruments and qualified oversight. A review of a relational construct can identify these research questions without providing individual medical recommendations. Claims about treatment efficacy, diagnostic equivalence, or medication decisions remain subject to specialist clinical evidence and independent medical review beyond the present nonclinical synthesis.
Function-Allocation Measurement and Construct Validity
Construct validity requires connecting the proposed term to observable regularities and a defensible theory of what the measurements represent. Cronbach and Meehl framed validation as the development and testing of a nomological network, while Borsboom and colleagues analyzed the relationship between an attribute and measurements intended to capture it. For relational redistribution, the first step is to show that distinct functions can be coded reliably and that a change in allocation can be separated from overall intensity of AI use, perceived companionship, and generic support seeking. A respondent's agreement with the statement 'AI has changed my relationships' is a useful perception but an inadequate standalone measure of the function-level construct. Cronbach & Meehl (1955) Borsboom et al. (2004)
A function-allocation matrix should cross time, function, recipient, provider, and performance mode. Rows might contain reassurance, interpretation, remembering, rehearsal, planning, and instrumental coordination; columns might identify self, specific human partners, artificial systems, organizations, and combinations of providers. Each observation should record whether the source supplied the function independently, suggested a step later performed by someone else, or jointly contributed to the result. Separate quantities capture episode frequency, time spent, subjective utility, perceived reliability, and later outcomes. A person-level concentration score can be calculated from these observations, although its validity must be tested rather than inferred from mathematical convenience.
Validation proceeds through cognitive interviews to learn what participants mean by each category, interrater agreement on event descriptions, convergent comparisons with task or log measures, and discriminant tests against neighboring constructs. Measurement invariance across human and artificial provider labels, languages, ages, and cultural contexts should be examined before comparing scores. Test-retest reliability is meaningful only under conditions where the distribution itself is expected to remain stable; genuine life transitions should produce change. The investigator should also assess whether technical logs systematically undercount offline human support and whether retrospective surveys inflate unusually memorable AI exchanges. These checks determine whether observed migration reflects psychology or a change in measurement visibility.
Longitudinal, Multilevel, and Causal Research Designs
A credible longitudinal study begins with a clearly specified change process. Let each function at a given time have an observed set of providers, along with contextual measures of demand, access, and satisfaction. The key dependent variables are subsequent allocations and downstream outcomes such as interpersonal communication, stress appraisal, relationship quality, and task success. Within-person analyses can estimate whether deviations from a person's usual AI access coincide with shifts in a particular function. They must also represent reciprocal feedback: an earlier conflict can increase help seeking, while earlier help seeking may influence the next conflict. Observational time ordering strengthens description but does not by itself remove confounding.
Experimental approaches can manipulate which resources are available, the type of response provided, the timing of access, or whether a technical system can retrieve prior context. A factorial design can isolate effects of informational accuracy and social framing; a staged rollout can create a comparison around a service feature, subject to the assumptions of the rollout mechanism. Appropriate controls should receive meaningful alternative assistance rather than artificial deprivation when emotional needs are involved. The primary outcome must correspond to the function under test. An intervention designed to improve conflict rehearsal should measure subsequent communication rather than relying exclusively on chatbot satisfaction or generalized loneliness scores.
Multilevel analysis is needed because episodes occur within people, relationships, households, workplaces, and platforms. One participant's allocation affects another person's workload and expectations; this interdependence violates naive assumptions of independent observations. Dyadic and network models may represent reciprocal influence, while time-varying platform policies enter as higher-level exposures. Consent and privacy restrictions shape what evidence can be collected from partners or private conversations. Preregistration should identify primary outcomes and causal contrasts; sensitivity analysis should examine unmeasured confounding, attrition, and classification uncertainty. These methodological commitments transform the idea of relational redistribution into a program of independently falsifiable tests. Bandura (2001) Lee & See (2004)
Cultural, Developmental, and Institutional Moderators
Relational functions are defined and evaluated within social expectations. Families, friendship circles, workplaces, and professional services differ in who is expected to provide emotional reassurance, practical aid, memory, and conflict interpretation. A change in the allocation of a function may therefore have different meanings in different cultural and institutional environments. For example, consulting an outside adviser before a family discussion can be understood as ordinary preparation in one context and a breach of intimate boundaries in another. Research should document local norms of disclosure, reciprocity, obligation, privacy, and acceptable mediation rather than assume a universal reaction to AI assistance.
Developmental context also matters. Adolescents, emerging adults, parents, caregivers, and older adults may face different social tasks, support networks, technology access, and patterns of reliance on others. A study based on university participants cannot automatically represent families with dependent children or older adults experiencing bereavement. The 2026 peer-versus-chatbot experiment among students and the naturalistic Character.AI survey address different populations and designs; their apparent differences may reflect both circumstances and intervention structure. Representative sampling and stratified analysis are needed before population-wide estimates of migration or risk can be proposed. Research on emotional support should also attend to disability, language, economic resources, and differences in access to trustworthy human services. Li et al. (2026) Zhang et al. (2026)
Institutional moderation concerns the rules through which resources become available and governable. An employer may require employees to use AI to draft difficult correspondence, a university may offer counseling-related tools, or a platform may change personalization and retention features. Such arrangements alter practical agency, the confidentiality of exchanges, and the ability to refuse delegated functions. Relational redistribution may therefore reflect organizational power rather than individual preference alone. Studies should separate personally chosen AI use from imposed or default-mediated use, and consider whose workload or information access changes. An adequately specified configuration includes people, technologies, institutions, and the social norms through which their activities acquire meaning.
Bias, Measurement Failure, and Replication
Selection bias is a central problem in the present evidence. People seeking artificial companionship may already have distinctive levels of loneliness, anxiety, interest in technology, or availability of human support. Survey studies can describe these patterns but generally cannot identify how the same individual would have fared without access. Even longitudinal associations may be shaped by changing life circumstances and reciprocal decisions. The studies by Folk and Dunn and Zhang and colleagues should accordingly be read as evidence of patterned associations and temporal relationships within their designs, alongside distinct randomized evidence from short-term interventions. Convergent findings across designs warrant attention; apparently conflicting results require analysis of outcome definitions, samples, and exposure conditions. Folk & Dunn (2026) Zhang et al. (2026)
Measurement error may arise when researchers infer relational function from message text alone. Phrases such as 'thank you,' 'I trust you,' or 'I need you' can occur during humor, creative role-play, practical work, distress, or attachment-related exchange. Automated classifiers require independent human validation, category-specific error reporting, and attention to language and cultural context. Platform logs can miss human contact and cannot directly measure the motives of someone absent from the conversation. Self-reports can suffer recall bias and social desirability effects. Triangulating user diaries, optional partner reports, verified logs, and task-level outcomes provides a stronger basis for claims about allocation than any single source.
Replication requires specifying the service configuration and preserving enough information for independent reconstruction without exposing private conversations. Relevant metadata include model and application version, context and memory settings, response policies, user population, recruitment pathway, task instructions, and preregistered outcomes. Investigators should report how missing data were handled and whether analyses were developed after inspecting results. Repeated studies ought to vary product type and social setting to reveal boundary conditions. A result that appears under one companionship interface and one temporary policy may represent a real contextual effect while failing to generalize to other systems. Scientific progress depends on mapping those conditions rather than seeking a universal average detached from actual relational practices.
Safety, Governance, and Distributional Consequences
Redistribution changes who can influence interpretation, memory, and decisions. When a system becomes a regular intermediary, its designers and operators acquire a role in shaping the conversational environment. Privacy settings, data-retention periods, ability to export histories, model updates, and limitations on sensitive advice become experimentally relevant institutional conditions. A safety assessment should evaluate the accuracy of consequential outputs, the handling of third-party information, and whether users retain opportunities to verify assumptions with affected people. These issues connect technical governance to psychological outcomes such as perceived control, continuity, conflict, and trust. A study of the user's satisfaction with a response alone cannot describe this institutional distribution of influence.
The loss or alteration of AI companionship can produce distress for some users, as documented in 2026 research on companion disruption. A research and governance response should identify whether distress follows the disappearance of a perceived social bond, the loss of archived memory, the interruption of a daily coping routine, or the combination of several functions. Each pathway suggests different measurements and possible mitigation strategies. Portability, advance notice of major service changes, access to records, and available human support are variables worth evaluating, rather than assumed solutions with established effect sizes. Interventions must be assessed for unintended consequences, including increased exposure of sensitive information or false promises of continuity. De Freitas et al. (2026)
Distributional effects extend beyond the person using the system. Delegating planning may reduce one partner's workload but increase another's verification responsibilities; confidential rehearsal may improve a difficult conversation but exclude a partner from opportunities for mutual interpretation. Institutions may benefit from automation while shifting the risks of errors onto clients or workers. A complete account identifies the beneficiaries, burden bearers, and decision rights associated with each transferred function. It should examine fairness across users with different literacy, language access, disability status, and resources. Responsibility remains traceable through actual human and institutional decision structures, regardless of whether participants experience the artificial interface as socially engaging.
Integrated Synthesis and Research Agenda
The literature supports five grounded propositions for organizing subsequent investigation. First, social support is functionally diverse, and its effects depend on available resources and the conditions of use. Second, cognition and remembering can involve external representational systems whose role is demonstrated through task-specific behavior. Third, human social responses to computing interfaces vary with framing and attribution. Fourth, naturalistic AI companionship is associated with heterogeneous experiences and psychosocial outcomes, and controlled studies produce results tied to intervention, comparator, population, and duration. Fifth, platform changes can disrupt routines and relationships that users have organized around persistent services. Each proposition is supported by a distinct body of work; their integration makes the proposed construct scientifically useful while leaving its distinctive predictions to be tested. Cohen & Wills (1985) Hutchins (1995) Nass & Moon (2000) Li et al. (2026) De Freitas et al. (2026)
The first research priority is a validated taxonomy of relational functions. Investigators should use cognitive interviews and real event records to test whether reassurance, interpretation, remembering, rehearsal, and planning are sufficiently separable, and whether additional functions emerge in different populations. The second priority is a repeated allocation instrument with time-stamped providers, multiple-provider coding, task demand, and independent checks on behavioral consequences. Third, comparative experiments should test whether system accuracy, perceived responsiveness, memory access, and governance rules have distinct effects on function assignment. Fourth, longitudinal dyadic and social-network studies should examine whether changes precede or follow relational outcomes while accounting for reciprocal influence.
A fifth priority is cross-platform replication with transparent version histories and meaningful negative results. Studies should examine discontinuations, access restrictions, and migrations between systems, since such events reveal which functions people have come to locate in specific resources. Sixth, researchers should model distributions of responsibility and privacy burdens among users, third parties, and institutions. The strongest evidence would permit investigators to distinguish beneficial delegation, temporary coping, social displacement, and function-preserving reorganization within the same individual over time. This framework specifies observable relationships, falsifiable contrasts, and the interpretive conditions necessary for evidence to accumulate. It is an open research agenda, not a report of experiments performed for this monograph.
Conclusion
Relational Function Redistribution is a proposed, operationally tractable description of how psychological and communicative functions are reassigned, shared, and transformed among humans, artificial systems, and institutions through time. Its central scientific claim is that relationship change may occur at the level of functions even when the number of human contacts remains constant. Reassurance, interpretation, autobiographical remembering, conflict rehearsal, planning, and support can be traced as activities with identifiable providers, purposes, and consequences. Their allocation should be measured as a dynamic configuration, with attention to task demand and the possibility that several sources contribute to the same event. A function-level lens provides a specific way to connect social psychological and cognitive research without erasing their separate mechanisms.
Established literatures provide foundations for this approach. Social support studies identify different pathways through which relationships can matter for adjustment; transactive memory and cognitive offloading demonstrate the role of external sources in remembering and coordination; human–computer interaction experiments reveal systematic social responses to interfaces; and contemporary AI companionship research documents diverse patterns of use, subjective significance, and psychosocial association. Experimental work offers more narrowly supported causal findings for particular interventions and comparators. Available evidence does not yet establish a population-wide causal effect of the named redistribution construct, nor a single direction of consequence across contexts. That limitation specifies the need for longitudinal and experimentally informed measurement rather than undermining the value of a well-defined research object. Feeney & Collins (2015) Sparrow et al. (2011) Epley et al. (2007) Folk & Dunn (2026)
The resulting scientific program brings three questions into focus: which function changes providers, what process explains the change, and what consequences unfold for the people and institutions involved? Answering them requires validated event coding, repeated function-allocation matrices, independent measures of outcomes, documented system changes, and evidence from several participants wherever appropriate and consensual. It also requires the separation of a researcher's theoretical proposal from the empirical claims of the studies used to develop it. Within the Artificial Era, these methods make it possible to examine human–artificial configurations as changing organizations of psychological work, while retaining the precision and accountability demanded by psychological science.
References
- Bandura, A. (2001). Social cognitive theory: An agentic perspective. Annual Review of Psychology, 52, 1–26. https://doi.org/10.1146/annurev.psych.52.1.1
- Borsboom, D., Mellenbergh, G. J., & van Heerden, J. (2004). The concept of validity. Psychological Review, 111(4), 1061–1071. https://doi.org/10.1037/0033-295X.111.4.1061
- Clark, A., & Chalmers, D. (1998). The extended mind. Analysis, 58(1), 7–19. https://doi.org/10.1093/analys/58.1.7
- Cohen, S., & Wills, T. A. (1985). Stress, social support, and the buffering hypothesis. Psychological Bulletin, 98(2), 310–357. https://doi.org/10.1037/0033-2909.98.2.310
- Cronbach, L. J., & Meehl, P. E. (1955). Construct validity in psychological tests. Psychological Bulletin, 52(4), 281–302. https://doi.org/10.1037/h0040957
- De Freitas, J., Castelo, N., Uğuralp, A. K., & Oğuz-Uğuralp, Z. (2026). Mourning the loss of AI companions. Nature Human Behaviour. Advance online publication. https://doi.org/10.1038/s41562-026-02569-3
- Epley, N., Waytz, A., & Cacioppo, J. T. (2007). On seeing human: A three-factor theory of anthropomorphism. Psychological Review, 114(4), 864–886. https://doi.org/10.1037/0033-295X.114.4.864
- Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., Maes, P., Phang, J., Lampe, M., Ahmad, L., & Agarwal, S. (2025). How AI and human behaviors shape psychosocial effects of extended chatbot use: A longitudinal controlled study [Preprint]. MIT Media Lab. https://www.media.mit.edu/publications/how-ai-and-human-behaviors-shape-psychosocial-effects-of-chatbot-use-a-longitudinal-controlled-study/
- Feeney, B. C., & Collins, N. L. (2015). A new look at social support: A theoretical perspective on thriving through relationships. Personality and Social Psychology Review, 19(2), 113–147. https://doi.org/10.1177/1088868314544222
- Fitzpatrick, K. K., Darcy, A., & Vierhile, M. (2017). Delivering cognitive behavior therapy to young adults with symptoms of depression and anxiety using a fully automated conversational agent (Woebot): A randomized controlled trial. JMIR Mental Health, 4(2), e19. https://doi.org/10.2196/mental.7785
- Folk, D., & Dunn, E. (2026). How does turning to AI for companionship predict loneliness and vice versa? Psychological Science, 37(4), 276–286. https://doi.org/10.1177/09567976261427747
- Hazan, C., & Shaver, P. (1987). Romantic love conceptualized as an attachment process. Journal of Personality and Social Psychology, 52(3), 511–524. https://doi.org/10.1037/0022-3514.52.3.511
- Hollan, J., Hutchins, E., & Kirsh, D. (2000). Distributed cognition: Toward a new foundation for human-computer interaction research. ACM Transactions on Computer-Human Interaction, 7(2), 174–196. https://doi.org/10.1145/353485.353487
- Hutchins, E. (1995). Cognition in the wild. MIT Press. https://doi.org/10.7551/mitpress/1881.001.0001
- Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46(1), 50–80. https://doi.org/10.1518/hfes.46.1.50_30392
- Li, R.-N., Folk, D., Singh, A., Ungar, L., & Dunn, E. (2026). Is a random human peer better than a highly supportive chatbot in reducing loneliness over time? Journal of Experimental Social Psychology, 125, Article 104911. https://doi.org/10.1016/j.jesp.2026.104911
- Lucas, G. M., Gratch, J., King, A., & Morency, L.-P. (2014). It's only a computer: Virtual humans increase willingness to disclose. Computers in Human Behavior, 37, 94–100. https://doi.org/10.1016/j.chb.2014.04.043
- Nass, C., & Moon, Y. (2000). Machines and mindlessness: Social responses to computers. Journal of Social Issues, 56(1), 81–103. https://doi.org/10.1111/0022-4537.00153
- Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002
- Sparrow, B., Liu, J., & Wegner, D. M. (2011). Google effects on memory: Cognitive consequences of having information at our fingertips. Science, 333(6043), 776–778. https://doi.org/10.1126/science.1207745
- Wegner, D. M. (1987). Transactive memory: A contemporary analysis of the group mind. In B. Mullen & G. R. Goethals (Eds.), Theories of group behavior (pp. 185–208). Springer. https://doi.org/10.1007/978-1-4612-4634-3_9
- Zhang, Y., Zhao, D., Hancock, J. T., Kraut, R., & Yang, D. (2026). Interaction with AI companions and psychological well-being. Nature Human Behaviour, 10, 1966–1980. https://doi.org/10.1038/s41562-026-02516-2
Social Support as an Allocation of Resources
Social support research provides a foundational account of why the distribution of functions matters. Cohen and Wills distinguished processes through which supportive relationships might buffer adverse effects of stress and broader beneficial associations between social integration and well-being. The buffering hypothesis concerns the conditions under which a supportive resource changes the relationship between stress exposure and psychological outcomes. In relational redistribution research, the crucial questions are which support functions are available, whether they are perceived as credible, and whether the person can mobilize them when required. The same availability of conversational contact can sustain different kinds of support, depending on trust, prior history, response specificity, and practical action. Cohen & Wills (1985)
Feeney and Collins elaborated a perspective in which close relationships can support thriving both under adversity and through the facilitation of exploration and growth. This framework extends measurement beyond immediate relief. A system that supplies encouraging language might improve an isolated episode of subjective comfort, while a human partner who helps enact a difficult decision may contribute to sustained competence and connectedness. These are empirically separable pathways. Investigators should distinguish perceived support during interaction, enacted help, subsequent behavior, repeated relationship investment, and longer-term outcomes. The expectation that a provider will remain accessible may itself affect regulation before any advice is received. Feeney & Collins (2015)
Human support is embedded in obligations, reciprocity, social identity, and shared accountability. Artificial systems can generate informational responses and simulated empathic language, but the meaning of support depends on what the recipient believes the exchange accomplishes and what the system can reliably do. A technically fluent answer may be appraised as supportive when it provides a usable interpretation; the same answer can create miscalibrated certainty when sources are inaccurate. A resource-allocation account therefore needs separate indicators for availability, response quality, responsibility, reciprocity expectations, and downstream human behavior. It should allow participants to identify when the system served as a supplement, an intermediary, a rehearsal environment, or the main source of the relevant function.