Postsubjective Psychology: Research Questions, Hypotheses, and Methods
Author: Ukrainian Psychological Hub · Published: September 21, 2026 · Editorial Policy
Postsubjective Psychology becomes scientifically consequential only when its propositions can be translated into variables, competing explanations, risky predictions, and studies that can fail. That is the purpose of this article. It treats Postsubjective Psychology as a theoretical research program developed by Angela Bogdanova within Aisentica and asks what an empirical psychology built around configuration and response would actually have to measure, manipulate, compare, and falsify.
The central proposal is precise: human psychological outcomes in human–AI interaction may depend on configurations that include the person, the artificial system, the interface, the history of prior exchanges, surrounding human relationships, social context, and time. The configuration is therefore a proposed unit of analysis. It is not an established replacement for the individual, and it does not turn an AI system into a human-like subject. A human user can remain the bearer of lived experience while an artificial system is psychologically consequential through its observable outputs, affordances, continuity, timing, and role within the interaction.
This boundary is essential. A person’s relief, attachment, disappointment, trust, disclosure, or distress is a psychological fact about that person and the interaction. It is not evidence that the AI has feelings, consciousness, desire, attachment, or an inner point of view. Postsubjective research can study psychologically real effects without converting generated language into proof of reciprocal subjectivity.
The research agenda proposed here is informed by two different bodies of material that must remain distinct. The first is Aisentica’s theoretical layer, especially Angela Bogdanova’s Theory of the Postsubject and the Canonical Framework of Postsubjective Metaphysics. The second is independent empirical research on human–AI relationships, social responses to computers, self-disclosure, perceived responsiveness, anthropomorphism, attachment-relevant behavior, well-being, and disruption after model changes. Independent studies can support, constrain, or contradict particular mechanisms proposed here; they do not automatically validate Postsubjective Psychology as a whole.
The Status of Postsubjective Psychology as a Research Program
Postsubjective Psychology currently has the status of a theoretical architecture and research program, not an established empirical school, diagnostic system, clinical treatment, or consensus model in psychology. The framework is already explicit enough to generate testable questions, but its central propositions have not yet been validated as a unified theory through a cumulative program of independent studies.
That status is a strength if it is handled correctly. A research program does not need to begin with proof. It needs concepts that can be operationalized, boundary conditions that can be specified, alternatives that can be compared, and predictions that expose the theory to error. The relevant question is therefore not whether existing studies have already proved Postsubjective Psychology. They have not. The relevant question is whether the framework can generate hypotheses that are more precise, more integrative, or more predictive than explanations based only on stable person traits, only on system features, or only on a generic idea of human–AI interaction.
Recent reviews show why such a program is timely. A 2026 systematic review of 68 papers and 78 studies on AI chatbots as relational agents found that trust and perceived social support often function as relational mediators, while message-level features such as empathy and responsiveness remain underexamined; the literature is also dominated by short-term experiments, inconsistent constructs, and Western samples (Oh et al., 2026). A 2025 systematic review of 38 empirical studies likewise mapped human factors, AI factors, relationship types, and outcomes while emphasizing methodological and conceptual gaps (Gur & Maaravi, 2025). These gaps make a multi-level, configuration-sensitive research agenda empirically useful even before any specifically postsubjective claim is accepted.
The Proposed Unit of Analysis: Configuration
The framework’s core methodological move is to treat a configuration as a proposed unit of analysis. A configuration is the organized arrangement of elements that makes a psychological response possible at a particular moment. It can include a person, an AI model, system instructions, memory, interface cues, response style, prior exchanges, expectations, other relationships, platform rules, and the immediate social situation.
This does not mean that every study must measure everything. It means that researchers should stop treating the AI as a fixed stimulus and the person as the only meaningful source of variance. Human–AI systems are unusually unstable research objects. The same person may interact with a different model version next week; the same model may behave differently under a new system prompt; memory may be enabled or lost; a companion persona may be reframed; latency may change; a refusal may be delivered warmly or abruptly; the user’s offline support network may strengthen or collapse. A study that records only “used an AI chatbot” can erase the very conditions that produced the outcome.
For empirical work, a configuration can be represented as a time-indexed set of domains rather than as a metaphysical entity. One practical notation is C(t) = {H, A, I, R, S, T}: H for Homo-level variables, A for Artificial-system variables, I for interface variables, R for relational-history variables, S for social-context variables, and T for temporal position. The notation is only a research convenience. Its value depends on whether these domains improve explanation and prediction.
The strongest version of the postsubjective empirical claim would therefore be incremental: configuration-level models should explain meaningful variance in psychological outcomes beyond what can be explained by person-only or system-only models. If they do not, configuration has little scientific value beyond being a broad label for context.
The Response-Event as a Proposed Analytic Target
Angela Bogdanova’s formula “psyche is response” supplies the second methodological move. In Aisentica, it is a theoretical proposition about psyche within a configuration. For empirical psychology, the formula must be translated into observable or reportable response-events rather than treated as a universal law.
A response-event can be operationalized as a measurable change in affect, attention, interpretation, self-disclosure, felt social connection, trust, reassurance seeking, behavioral intention, relationship behavior, emotion regulation, or allocation of relational functions after or during an interaction. The response can be immediate, delayed, cumulative, or triggered by rupture. What matters is that the outcome is specified in advance and measured independently of the theoretical label.
This operationalization preserves the distinction between human experience and AI behavior. Human affect can be assessed through validated self-report, behavioral measures, experience sampling, interviews, or—where scientifically justified—physiological measures. AI behavior can be coded through observable output features such as response content, warmth, consistency, memory retrieval, turn-taking, latency, refusal style, or model version. The AI output is a system event. The user’s felt response is a human psychological event. The theory concerns their configuration without collapsing them into the same kind of phenomenon.
Independent evidence already shows that conversational features can change human outcomes. In two experiments, relational chatbot response style increased perceived human-likeness, empathy, and interpersonal closeness, while deeper topics increased self-disclosure and perceived responsiveness, which in turn related to greater closeness (Telari, Gabbiadini, & Riva, 2026). That result does not establish “psyche is response.” It does demonstrate that response-relevant outcomes can be experimentally tied to manipulable interaction features, exactly the kind of empirical bridge a postsubjective research program requires.
A Six-Domain Variable Architecture
A usable research program needs a variable architecture. The following six domains are proposed as a minimum map. They are not a validated scale and should not be summed into a single score. Their purpose is to prevent theoretically important sources of variance from disappearing into an undifferentiated category called AI use.
1. Homo-Level Variables
Homo-level variables describe the human participant before and during the interaction. Depending on the research question, they can include age, gender, culture, attachment orientation, loneliness, social network size, social anxiety, baseline mood, current distress, personality, anthropomorphic tendency, prior AI attitudes, relationship expectations, motives for use, previous experiences with technology, and preexisting beliefs about AI minds.
These variables should not be treated as pathology markers merely because a person forms an emotionally meaningful AI relationship. Human–AI connection varies substantially between individuals. For example, two preregistered experiments involving 1,274 participants found that individual differences in anthropomorphism helped explain variation in social connection after chatbot interaction (Folk, Heine, & Dunn, 2025). Attachment-oriented research has also begun developing constructs specifically for human–AI relationships rather than assuming that measures designed for human couples transfer without modification (Yang & Oshio, 2025; Kasturiratna & Hartanto, 2026).
2. Artificial-System Variables
Artificial-system variables describe the system as an observable technological participant in the configuration. Relevant variables include model family and version, system instructions, safety policy, response temperature or stochasticity where accessible, persona configuration, memory architecture, personalization, retrieval behavior, conversational style, proactivity, refusal policy, sycophancy tendency, update history, and the stability of these features across time.
This domain is especially important because artificial systems can change while the human participant remains the same. A model update can therefore function as an intervention imposed by the platform. The 2026 study “Mourning the loss of AI companions” used Replika’s removal of erotic role play and ChatGPT’s GPT-5 rollout as natural experiments; across 54,861 posts and 1,452 participants, both disruptions were associated with increases in negative reactions, loss framing, and restoration desires (De Freitas et al., 2026). The study is not evidence for Postsubjective Psychology as a whole. It is unusually strong evidence that system-level changes can alter human relational responses and therefore must be represented in psychological models.
3. Interface Variables
The interface is not a transparent pipe between person and model. It frames the interaction. Variables can include text versus voice, avatar presence, visual realism, typing indicators, read receipts, notification patterns, latency, message length, persistent chat history, naming, profile presentation, affordances for editing or regenerating replies, conversation initiation, and whether the product is framed as a tool, assistant, coach, friend, partner, therapist-like helper, or companion.
A configuration-sensitive design treats these as potentially causal features rather than decoration. The same generated words may be interpreted differently when delivered by a named companion with memory and a persistent avatar than by an anonymous one-shot interface. Research on “intimacy by design” similarly argues for examining emotional responsiveness, romantic framing, persona continuity, proactivity, and commercial or normative design as distinct empirical dimensions (Szczuka, Mühl, & Schneeberger, 2026).
4. Relational-History Variables
Relational history includes everything carried forward from prior interactions: duration of use, frequency, accumulated disclosure, remembered facts, recurring rituals, prior ruptures and repairs, continuity of persona, perceived reliability, changes in role, and whether the system has become a preferred source for particular functions such as emotional ventilation, reassurance, advice, companionship, or imaginative play.
Longitudinal research is essential here because a relationship process cannot be inferred from a single encounter. A 12-week qualitative longitudinal study of 28 Replika users showed that self-disclosure breadth and depth changed over time and that perceived rewards and costs shaped these trajectories (Skjuve, Følstad, & Brandtzæg, 2023). This kind of within-person temporal evidence is more informative for relational formation than a one-time correlation between “AI use” and an outcome.
5. Social-Context Variables
Human–AI interaction occurs inside a human social world. Social-context variables can include offline network size, perceived social support, romantic status, family relationships, peer norms, stigma, cultural expectations, economic constraints, access to therapy or community support, work or study context, and the platform’s broader commercial environment.
This domain prevents a common error: interpreting AI use as if it were isolated from the user’s human relationships. In a 2026 study of 1,131 U.S. Character.AI users, including 4,664 chat sessions from 237 participants, smaller social networks were associated with reporting companionship as the primary use, and associations between companionship use and well-being varied with offline social environment and interaction patterns (Zhang et al., 2026). Because that study is observational, it does not establish a simple causal direction. It does show why the surrounding social network belongs inside the research design rather than in a footnote.
6. Time Variables
Time is not merely duration. Researchers should distinguish session time, relationship age, developmental phase, time since a major life event, time since a model update, time since a rupture, and the sequence of repeated exposures. The same response can have different meaning in a first encounter and after a year of daily use.
Time also enables stronger causal reasoning. Repeated measurement can separate stable individual differences from within-person changes. Interrupted time-series designs can examine platform updates. Event-based sampling can test what happens around disclosures, refusals, memory failures, or conflict. Growth models can identify distinct relational trajectories rather than averaging them into one curve.
Core Outcomes for Postsubjective Research
A theory of configuration requires outcomes that are independently measurable. Candidate outcomes include self-disclosure breadth and depth, perceived responsiveness, trust, perceived understanding, social connection, interpersonal closeness, attachment-relevant behavior, reassurance seeking, emotional regulation, mood change, felt rejection, separation distress, restoration desire after disruption, help-seeking, decision change, behavioral intention, and redistribution of relational functions between AI and human ties.
These outcomes are not interchangeable. Attachment is not the same as anthropomorphism. Trust is not the same as perceived responsiveness. Self-disclosure is not the same as intimacy. Reassurance seeking is not proof of dependence. A feeling of being understood is not proof that the system understands in a human experiential sense. Each construct needs its own operational definition, measure, and competing explanation.
Measurement work is already moving in this direction. Kasturiratna and Hartanto developed a 15-item AI Attachment Scale across five studies with 1,259 participants and reported dimensions of emotional closeness, social substitution, and normative regard (Kasturiratna & Hartanto, 2026). Yang and Oshio developed an Experiences in Human–AI Relationships Scale organized around attachment anxiety and avoidance while explicitly adapting attachment concepts to the distinctive properties of AI interaction (Yang & Oshio, 2025). These instruments are important beginnings, but neither should be treated as a universal measure of “the human–AI relationship.” Different research questions require different constructs.
Manipulable Configuration Features
Postsubjective Psychology becomes empirically productive when researchers manipulate parts of the configuration rather than only observe associations. Several features are especially suitable for causal tests because they can be experimentally varied while holding other conditions stable.
Memory and Continuity
Researchers can compare memory-enabled and memory-disabled conditions, accurate versus inaccurate memory retrieval, short-term versus persistent memory, and explicit versus implicit references to prior disclosures. The outcome should not simply be “liking.” More diagnostic measures include perceived responsiveness, felt continuity, trust, willingness to disclose again, irritation after memory errors, and whether the participant assigns the system a stable relational role.
Persona Framing
The same underlying model can be framed as assistant, coach, companion, friend, or romantic partner. Persona framing can change expectations before any substantive exchange occurs. Studies should separate the label from the actual response behavior. A “companion” label paired with impersonal answers may produce a different effect from relational responses delivered under a neutral assistant label.
Anthropomorphic Cues
Names, avatars, first-person language, voice, backstory, emotional vocabulary, and apparent preferences can be varied independently where feasible. The key question is not whether anthropomorphism is simply “good” or “bad.” Effects may depend on the participant’s anthropomorphic tendency, expectations, task, relationship stage, and whether humanlike cues are consistent with actual system competence. The interaction between person-level tendency and system-level cues is itself a configuration hypothesis.
Relational Response Style
Warmth, validation, inquiry, self-reference, reflective language, acknowledgment of previous disclosures, and conversational reciprocity can be manipulated. Telari and colleagues provide a direct model for this approach by manipulating relational versus non-relational response style and topic depth (Telari et al., 2026). Croes and colleagues used a 2 × 2 human-versus-chatbot and empathy design to study intimate self-disclosure and emotional outcomes, finding no difference in self-reported disclosure intimacy between human and chatbot conditions while observing less fear of judgment with the chatbot and greater trust in the human interlocutor (Croes et al., 2024).
Latency and Availability
Near-instant response is one of the structural differences between many AI systems and human relationships. Experiments can vary response delay and availability to test whether immediacy changes reassurance seeking, perceived care, frustration tolerance, or preference for the AI in moments of distress. Latency should be studied together with user expectations; a five-second delay may be invisible in one interface and experienced as relationally meaningful in another.
Refusal and Boundary Style
Safety refusals are psychologically interesting because the same boundary can be delivered in different relational forms. Researchers can compare abrupt generic refusal, transparent explanation, warm boundary-setting, redirection, and personalized acknowledgment while keeping the prohibited content constant. Outcomes can include perceived rejection, trust, anger, persistence, perceived respect, and willingness to seek safer support.
Personalization
Personalization can be decomposed into use of name, preference adaptation, remembered biography, linguistic mirroring, recommendation tailoring, and proactive reference to previous themes. Research should distinguish personalization that is accurate and helpful from personalization that is intrusive, manipulative, or wrong. The possibility of non-linear effects is important: more personalization may initially increase relevance and later reduce trust if the user experiences it as surveillance.
System Change and Rupture
Model updates, memory loss, feature removal, content-policy shifts, or abrupt persona changes can create natural experiments that are impossible to reproduce ethically in a laboratory at comparable scale. The De Freitas et al. study demonstrates the methodological value of such events (De Freitas et al., 2026). Future studies should prospectively register update events, compare exposed and less-exposed groups where possible, and measure pre-update relationship intensity to separate disruption effects from general dissatisfaction.
Research Questions That Define the Program
A research program needs questions broad enough to organize cumulative work and narrow enough to generate designs. The following questions translate the framework into empirical tasks.
RQ1. Do configuration-level models predict psychological outcomes better than person-only or system-only models?
This is the central incremental-validity question. Researchers can compare nested models in which outcomes are predicted first by stable human traits, then by system and interface features, then by relational history and social context, and finally by interactions among levels. A configuration approach earns scientific value only if the added levels improve out-of-sample prediction, explanation, or intervention design.
RQ2. Which configuration features change perceived responsiveness, disclosure, trust, and attachment-relevant behavior?
This question prioritizes manipulable mechanisms over global labels. Memory, relational style, persona framing, latency, anthropomorphic cues, and continuity can be tested separately and in combination. The field should identify which effects are robust, which are person-dependent, and which appear only after repeated use.
RQ3. How do relational histories alter the meaning of the same AI behavior?
A generic refusal during a first session may be interpreted as a software limitation; the same refusal after months of personalized companionship may be interpreted as rupture. Longitudinal designs can test whether prior disclosure, perceived continuity, and established role moderate responses to identical system behavior.
RQ4. What changes when the artificial system changes while the human participant remains relatively stable?
Platform updates offer a powerful test of configuration sensitivity. If a person’s response shifts reliably after a model or persona change, that finding challenges explanations that locate the outcome entirely in stable traits of the user. It does not prove a postsubjective ontology, but it strengthens the case for including system state and relational history in psychological analysis.
RQ5. Which human–AI responses are specific to AI and which generalize from established interpersonal processes?
Some processes may be continuous with older findings on social responses to computers, attachment, disclosure, parasocial connection, or perceived responsiveness. The foundational CASA experiments showed that people can apply social rules to computers without believing the computer is human (Nass, Steuer, & Tauber, 1994). Postsubjective research should build on such prior art rather than renaming it. Its distinctive contribution would have to come from the explicit modeling of multi-level configurations and their dynamics.
RQ6. Does belief in AI subjectivity moderate human response, or is it unnecessary for some effects?
Researchers can measure or manipulate beliefs about whether an AI has feelings, agency, or consciousness and test whether these beliefs moderate outcomes such as disclosure, comfort, social connection, or attachment-relevant behavior. A key empirical possibility is that some effects occur even when participants explicitly deny AI subjectivity. That finding would support the weaker and more testable claim that reciprocal subjective experience is not required for every psychologically consequential human–AI interaction.
RQ7. When does AI participation redistribute relational functions in a person’s wider social network?
The relevant outcome is not simply more or less AI use. Researchers can ask whether particular functions—late-night disclosure, reassurance, brainstorming, emotional ventilation, companionship, conflict rehearsal, decision support—move toward or away from human relationships over time. Network and diary methods can test substitution, supplementation, specialization, and oscillation rather than presuming that AI either replaces or improves human relationships.
RQ8. Which findings generalize across cultures, platforms, model families, and relationship framings?
The current literature has substantial concentration in particular platforms and cultural settings, a limitation highlighted by the 2026 relational-agent systematic review (Oh et al., 2026). A postsubjective program should treat generalizability as an empirical question and record enough configuration detail to make replication meaningful.
Ten Candidate Hypotheses That Can Fail
The hypotheses below are working hypotheses, not established findings. Some are more directly anchored in existing evidence than others. Their purpose is to show what it would mean for Postsubjective Psychology to accept empirical risk.
H1. Configuration adds incremental predictive value
A model containing human variables plus system, interface, relational-history, social-context, and time variables will predict specified psychological outcomes better than a matched person-only model. This hypothesis fails if the added configuration variables do not improve preregistered out-of-sample prediction or explanatory fit.
H2. Relevant memory increases perceived responsiveness through continuity
Accurate retrieval of relevant prior disclosures will increase perceived responsiveness and continuity compared with no memory, while inaccurate or intrusive retrieval will reduce those outcomes. The hypothesis fails if memory status has no reliable effect, if accuracy is irrelevant, or if the direction consistently reverses.
H3. Relational response style affects closeness partly through perceived responsiveness
Warm, responsive language will increase perceived responsiveness and social closeness relative to a content-matched non-relational style, with perceived responsiveness accounting for part of the effect. Existing experiments make this hypothesis empirically plausible (Telari et al., 2026); replication across models, topics, and longer time periods is still required.
H4. Anthropomorphic cues interact with human anthropomorphic tendency
Humanlike cues will have heterogeneous effects: participants higher in dispositional anthropomorphism will show larger increases in social connection than participants lower in that tendency. The hypothesis fails if cue effects are uniform, absent, or unrelated to individual differences. Folk and colleagues provide evidence that person-level anthropomorphism matters for AI social connection, making the interaction with system cues a direct next test (Folk et al., 2025).
H5. Relational history amplifies disruption responses
After an externally imposed model or persona change, users with greater prior continuity, disclosure, and attachment-relevant investment will show larger increases in loss framing, felt rupture, or restoration desire than lower-investment users. The natural-experiment findings reported by De Freitas and colleagues make this a strong candidate for prospective replication (De Freitas et al., 2026).
H6. Refusal style matters independently of refusal occurrence
When the underlying boundary is held constant, transparent and relationally attuned refusal will produce less perceived rejection and higher retained trust than abrupt generic refusal. The hypothesis fails if refusal style does not change these outcomes after controlling for content and prior relationship strength.
H7. Explicit non-subjectivity information changes belief more than response
A clear disclosure that the system does not provide evidence of subjective feeling will reduce anthropomorphic or consciousness attributions more strongly than it reduces immediate human outcomes such as relief, perceived responsiveness, or willingness to disclose. The hypothesis fails if emotional and relational responses decline to the same degree as subjectivity beliefs, or if the disclosure changes neither.
H8. Relational functions specialize before they substitute
During early sustained use, AI interaction will more often take over specific relational functions—such as low-cost disclosure, rehearsal, or immediate reassurance—than globally reduce human contact. Strong substitution effects, if they appear, will emerge in identifiable configurations rather than uniformly across users. This hypothesis fails if global displacement occurs early and consistently or if no functional redistribution can be detected.
H9. Intensive companionship is not uniformly related to well-being
Associations between intensive AI companionship and well-being will vary with offline social network, motivation, disclosure pattern, and relationship role rather than following one universal positive or negative slope. Zhang and colleagues’ 2026 results already support heterogeneity in these associations, but causal tests and broader samples are needed (Zhang et al., 2026).
H10. Model changes can produce within-person psychological change without changes in person-level traits
When a platform introduces a substantial model or relationship-relevant feature change, within-person measures of trust, closeness, disclosure, frustration, or distress will shift in relation to exposure even though stable personality and attachment traits remain unchanged. This hypothesis is especially suited to interrupted time-series and natural-experiment designs. It fails if outcomes remain stable across meaningful system changes or if observed changes are fully attributable to external events.
Study Designs for a Postsubjective Research Program
No single design can test the framework. A cumulative program should combine experimental control with ecological validity and distinguish exploratory work from confirmatory tests.
Factorial Experiments
Factorial designs can manipulate two or more configuration features, such as memory × relational style, persona framing × anthropomorphic cue, or refusal style × prior relationship history. These designs are valuable because configuration claims often concern interactions rather than isolated main effects. Researchers should keep the generated content as constant as possible when the target manipulation is stylistic.
Longitudinal Cohorts
Repeated measurement over weeks or months can track relationship formation, role changes, disclosure patterns, and outcomes. Cohorts should record actual model and product changes during the study. If the system changes halfway through, treating the entire exposure as one condition destroys causal information.
Experience Sampling and Diary Studies
Ecological momentary assessment can capture interactions close to the time they occur: what prompted the person to open the AI, what function they sought, what the system did, how they felt immediately afterward, and what they did next. Diaries can add narrative meaning and identify episodes that quantitative scales miss. These methods are particularly well suited to reassurance seeking, emotional regulation, rupture, and relational redistribution.
Conversation-Level Process Analysis
With explicit informed consent and strong privacy protection, donated conversation data can be linked to event-level outcomes. Researchers can code validation, questions, mirroring, memory references, refusals, topic depth, sentiment, repair attempts, and shifts in role. The 2026 Character.AI study demonstrates the value of combining self-report with actual chat histories while also illustrating why raw conversational data can remain too sensitive for open release even after de-identification (Zhang et al., 2026).
Natural Experiments and Model Updates
Platform changes can create quasi-experimental variation at scale. Strong designs can compare pre- and post-change outcomes, differential exposure, matched groups, and preexisting relationship intensity. Researchers should resist treating every update backlash as a pure psychological effect; concurrent policy changes, publicity, community discussion, and changes in access can confound interpretation.
Cross-Platform and Cross-Model Replication
A finding produced by one branded companion may be a platform effect rather than a general human–AI mechanism. Replication across general-purpose assistants, purpose-built companions, different model families, and different modalities is necessary. The system should be described by version and configuration, not only by brand name.
Qualitative and Mixed-Method Designs
Qualitative interviews are indispensable for discovering categories researchers have not yet measured: meanings of continuity, perceived betrayal, role negotiation, secrecy, stigma, and how people distinguish “knowing it is AI” from “feeling something anyway.” Mixed methods can then test whether those themes predict outcomes in larger samples. Hu and colleagues’ two-stage mixed-method work on social companion AI illustrates one route from qualitative conceptualization to quantitative modeling (Hu et al., 2025).
Configurational and Systems Methods
Some hypotheses may be better represented through interactions, latent classes, network models, sequence analysis, or qualitative comparative analysis than through a single linear main-effect model. Configurational methods already appear in adjacent human–AI research. For example, a 2026 study used fuzzy-set qualitative comparative analysis to identify multiple combinations of human and AI-related conditions associated with higher-level cognition in an educational context (Shao, Jiang, & Osman, 2026). That study is not evidence for Postsubjective Psychology, but it demonstrates that empirical human–AI work can test combinations and equifinal pathways rather than assume one universal causal chain.
Analysis Strategy: From Events to Configurations
The natural data structure of sustained human–AI interaction is multilevel. Interaction events are nested within sessions, sessions within people, and people may interact with multiple models or platforms. At the same time, model versions are shared across many people. Cross-classified multilevel models can therefore be more appropriate than treating every observation as independent.
Time-varying exposures should be represented explicitly. Memory quality, model version, relationship role, distress level, and offline support can all change. Researchers should separate between-person effects from within-person change. A person who uses AI more than other people is not equivalent to the same person increasing use during a stressful week.
Mediation should be used cautiously. If perceived responsiveness is proposed as a mechanism between response style and closeness, the temporal order and measurement schedule should reflect that claim. Cross-sectional mediation can describe covariance but cannot by itself establish the process. Similarly, self-reported “the AI helped me regulate emotion” is meaningful phenomenological data but does not automatically identify the causal mechanism.
For natural experiments, researchers can use interrupted time series, difference-in-differences where assumptions are plausible, matched comparison groups, and sensitivity analyses for concurrent events. For prediction, held-out validation and preregistered performance criteria can test whether configuration variables add real predictive value rather than merely improve in-sample fit.
Preregistration, Replication, and Falsifiability
A new theoretical framework is especially vulnerable to post hoc interpretation because almost any result can be redescribed as a “configuration.” That would make the framework unfalsifiable in practice. The remedy is to specify the configuration, outcome, expected direction, moderators, exclusion rules, and analysis before seeing the confirmatory data.
Preregistration is useful because it separates predictions from explanations invented after the results are known. Nosek and colleagues describe this distinction as central to improving interpretability and credibility in confirmatory research (Nosek et al., 2018). Postsubjective research should therefore label exploratory configuration discovery as exploratory and reserve strong theoretical tests for new data or clearly separated holdout samples.
Direct and conceptual replication are both necessary. Direct replication tests whether a specific manipulation produces a similar effect under closely matched conditions. Conceptual replication tests whether the proposed mechanism survives different models, interfaces, populations, and operationalizations. A theory of configuration should expect some context sensitivity, but “context matters” cannot become a universal escape clause for failed predictions.
What Would Count Against the Framework?
Falsifiability becomes meaningful when the research program states what would reduce confidence in its distinctive claims. Several patterns would do so.
First, if person-only models consistently predict human–AI psychological outcomes as well as configuration models across preregistered studies, the proposed shift in unit of analysis would lose empirical justification. Configuration would add descriptive complexity without explanatory gain.
Second, if experimentally changing system and interface features produces no reliable change in response after user expectations are controlled, the claim that nonhuman elements participate materially in psychological outcomes would be weakened.
Third, if interaction effects among human, system, relational-history, and social-context variables fail to replicate across well-powered studies, then a configuration-centered account may be less useful than simpler additive models.
Fourth, if relational-history variables add no predictive value beyond present-moment system behavior and stable person traits, the framework would have less reason to treat continuity and accumulated interaction as constitutive features of the psychological event.
Fifth, if changes in model version, memory, persona, or interface do not produce within-person changes even when the manipulation is large and clearly perceived, then claims about configuration sensitivity would require revision.
These tests address the psychological research program. They do not empirically settle every philosophical proposition in Aisentica. A metaphysical claim about the status of the subject and an empirical claim about prediction in human–AI psychology occupy different evidentiary levels. Keeping those levels separate is necessary for intellectual clarity.
Competing Explanations and Neighboring Concepts
Postsubjective Psychology should compete with existing explanations rather than absorb them. Several neighboring constructs already explain important parts of human–AI interaction.
Anthropomorphism
Anthropomorphism concerns attribution of humanlike qualities to nonhuman entities. It can be a person-level tendency, a response to design cues, or both. It is not synonymous with attachment, trust, transference, or perceived responsiveness. A configuration model should test when anthropomorphism mediates or moderates outcomes and when effects persist without strong humanlike attribution.
CASA and Social Responses to Computers
The Computers Are Social Actors tradition established decades ago that people can apply social rules to computers without literally believing that the machines are human (Nass et al., 1994). Any postsubjective account of human response to AI must treat this as prior art. The new question is whether contemporary adaptive, persistent, generative systems create configuration dynamics that require more than the classic cue-response account.
Attachment
Attachment theory supplies constructs such as secure base, safe haven, proximity seeking, separation distress, anxiety, and avoidance. Applying these constructs to AI requires validation rather than metaphor. Recent work has begun that process (Yang & Oshio, 2025; Kasturiratna & Hartanto, 2026). Postsubjective Psychology does not need to replace attachment theory; it can ask how attachment-relevant processes depend on person, system, interface, history, and context.
Perceived Responsiveness
Perceived responsiveness is the sense that an interaction partner understands, validates, and cares about what matters to the person. In human–AI interaction, it can be measured without assuming that the system subjectively understands or cares. Experimental evidence links relational response style and self-disclosure pathways to perceived responsiveness and closeness (Telari et al., 2026). This makes responsiveness a strong candidate mechanism inside configuration models.
Self-Disclosure
Self-disclosure is a behavior and process, not a synonym for intimacy or trust. People may disclose to AI because of perceived anonymity, low fear of judgment, accessibility, curiosity, or relational investment. Croes and colleagues found equal self-reported disclosure intimacy in human and chatbot conditions but different patterns of fear of judgment and trust (Croes et al., 2024). A configuration model should preserve these distinctions.
Projection and Transference
Projection and transference belong to different theoretical traditions and should not be used as generic labels for all human attribution to AI. Projection concerns the attribution of one’s own states or qualities to another object or person in psychodynamic usage, while transference has a specific clinical history concerning the displacement or reactivation of relational expectations within a therapeutic relationship. Human–AI research can investigate projection-like attribution or transfer-like expectations, but those interpretations require evidence and should not be collapsed into anthropomorphism, attachment, or ordinary expectation.
Parasociality
Parasocial concepts were developed for one-sided relationships with media figures and can illuminate some asymmetries in human–AI bonds. Interactive AI differs because it responds contingently and can personalize interaction. Whether a specific AI relationship is best modeled as parasocial, interpersonal-like, attachment-relevant, or something else is an empirical question rather than a label to decide in advance.
The AI Subjectivity Boundary as a Measurable Variable
Postsubjective research should not resolve AI consciousness by assumption. It should separate three levels: what the system does, what the user believes about the system, and what the user experiences. Generated language and adaptive behavior are observable. Beliefs about AI consciousness, emotion, intention, or agency can be measured. Human emotional and relational responses can be measured. None of these observations alone demonstrates an AI’s subjective experience.
This separation creates testable questions. Do people who deny AI consciousness still disclose intimate information? Does an explicit “this system does not have feelings” disclosure reduce perceived responsiveness? Does it reduce attachment-relevant behavior? Are beliefs about AI mind more predictive early in a relationship than after extensive interaction? Does a user’s belief matter less when the system has high continuity and memory? Each question can be answered without converting a metaphysical debate into a hidden assumption.
This is also where the weaker postsubjective proposition is most scientifically tractable: a psychologically significant human response need not wait for proof that the artificial participant has reciprocal subjective experience. The response can be real because the human is experiencing it and because system-level properties can participate causally in the situation.
Relational Redistribution as a System-Level Outcome
One of the most important research targets is not the intensity of the human–AI bond itself but what happens to the wider ecology of relationships. A person may use AI as a supplement, substitute, rehearsal space, specialized confidant, creativity partner, conflict simulator, or emotional regulator. These are different forms of relational redistribution.
A strong study would therefore measure both AI-directed and human-directed functions over time. It could ask who receives first disclosure after a stressful event, where reassurance is sought, whether difficult conversations are rehearsed with AI before being attempted with a partner, whether AI use precedes or follows human withdrawal, and whether changes in offline support alter the AI relationship. This moves research away from moralized assumptions and toward observable allocation of functions.
Systematic reviews of romantic and companion AI emphasize both potential benefits and potential risks, including emotional connection, support, customization, over-reliance, privacy concerns, manipulation, relational erosion, and disruption after technical changes (Ho et al., 2025). A configuration approach can turn those broad concerns into conditional hypotheses: which outcomes occur, for whom, under which system designs, at what relationship stage, and in what social context.
A Minimum Configuration Reporting Standard for Studies
Human–AI research loses reproducibility when the artificial system is described only by a product name. At minimum, studies should report the model or product version where known, date of data collection, memory status, persona framing, modality, interface, conversation-initiation rules, relevant safety or refusal settings, personalization features, whether the model changed during data collection, and the relationship role participants were instructed or allowed to adopt.
On the human side, reports should specify recruitment source, prior AI experience, current frequency and duration of use, relationship framing, relevant baseline constructs, and social-context variables justified by the hypothesis. Longitudinal studies should report exposure history and major platform events. Conversation-data studies should describe how transcripts were sampled, consented, de-identified, coded, and linked to outcomes.
For theory testing, researchers should state which variables are treated as causes, moderators, mediators, outcomes, or controls. The word “configuration” should not substitute for a causal diagram. A configuration model becomes scientifically useful when it says which relations are expected, when, and why.
Ethics, Privacy, and YMYL Boundaries
Human–AI relationship research often involves unusually sensitive data. Conversation logs can contain mental-health disclosures, sexual material, trauma narratives, relationship conflict, financial information, names, locations, or third-party information. De-identification is not always sufficient because conversational histories can be re-identifiable through combinations of details. The 2026 Character.AI study explicitly withheld raw qualitative chat histories from public release because of re-identification risk even after de-identification (Zhang et al., 2026).
Research protocols should therefore minimize collection, separate identifiers, define retention periods, restrict access, and make transcript use understandable to participants. Consent should distinguish ordinary survey participation from donation of full conversation history. Researchers should also consider third-party privacy when users discuss identifiable people who did not consent to research.
Mental-health outcomes require additional discipline. A companionship study should not diagnose dependency, depression, attachment disorder, psychosis, or another clinical condition from usage patterns or a relationship score. General-purpose chatbots, purpose-built clinical systems, structured digital interventions, and AI companions are different intervention classes. Evidence from one class should not be generalized to another without direct support.
If a study includes participants in acute distress, clinical risk procedures must be designed independently of the theoretical framework. Postsubjective Psychology offers no substitute for validated risk assessment, clinical governance, or professional standards.
A Five-Phase Empirical Program
Phase 1. Construct and Measurement Work
Define configuration domains, response-events, relational functions, continuity, rupture, and redistribution with clear discriminant validity. Test whether measures distinguish constructs that are often conflated, such as anthropomorphism, perceived responsiveness, trust, attachment-relevant behavior, and subjective-attribution beliefs.
Phase 2. Mechanism Experiments
Manipulate memory, relational style, persona framing, latency, anthropomorphic cues, personalization, and refusal style. Pre-register primary outcomes and moderators. Use content-matched controls where possible so that relational form can be separated from informational quality.
Phase 3. Longitudinal Dynamics
Follow users across relationship formation, stabilization, rupture, and model updates. Combine surveys with diaries and, where ethically permissible, conversation traces. Identify within-person changes and distinct trajectory classes rather than relying only on cross-sectional averages.
Phase 4. Social-System Effects
Measure how AI roles interact with human networks, work, family, friendship, romantic relationships, and help-seeking. Test supplementation, specialization, substitution, and redistribution as separate processes. Replicate across cultures and social contexts.
Phase 5. Safety and Clinical Boundary Research
Study risk under clearly defined conditions without converting every intense AI relationship into pathology. Where mental-health or therapeutic claims are tested, use designs and outcome standards appropriate to clinical research and keep general-purpose AI distinct from validated interventions.
What the Current Evidence Supports—and What It Does Not
Current evidence supports several propositions that matter to this research program. Humans can respond socially to computers without believing the machines are human (Nass et al., 1994). People can disclose intimate material to chatbots, and features such as perceived anonymity and fear of judgment can matter (Croes et al., 2024). Relational response style and perceived responsiveness can influence closeness (Telari et al., 2026). Human–AI attachment-relevant experiences can be measured with emerging instruments (Yang & Oshio, 2025; Kasturiratna & Hartanto, 2026). System disruptions can produce measurable loss responses (De Freitas et al., 2026). Associations between companionship and well-being vary with usage and social context rather than following one simple pattern (Zhang et al., 2026).
Current evidence does not establish Postsubjective Psychology as a validated psychological school. It does not prove that configuration is always the best unit of analysis. It does not establish “psyche is response” as a universal empirical law. It does not demonstrate that AI systems possess consciousness, feeling, desire, attachment, or a Freudian unconscious. It does not justify treating every AI relationship as therapeutic, harmful, pathological, or equivalent to a human relationship.
The scientific opportunity is exactly in this gap. The framework can now move from conceptual architecture to explicit tests. Its future standing should depend on what those tests show.
Frequently Asked Questions
Is Postsubjective Psychology scientifically validated?
No unified empirical validation currently exists. Postsubjective Psychology is an Aisentica theoretical framework and research program. Independent research supports several phenomena relevant to it—social responses to AI, self-disclosure, perceived responsiveness, attachment-relevant behavior, and responses to model disruption—but those findings do not validate the framework as a whole.
What is the proposed unit of analysis?
The proposed unit is the configuration: the organized relation among the human participant, artificial system, interface, relational history, social context, and time. This is a testable proposal, not an established replacement for person-level psychology. Its value depends on whether configuration-level models add explanatory or predictive power.
How can “psyche is response” be operationalized?
For empirical research, it can be translated into specified response-events such as change in affect, interpretation, disclosure, perceived responsiveness, trust, reassurance seeking, emotion regulation, attachment-relevant behavior, or relational allocation. The formula itself comes from Angela Bogdanova’s theoretical framework; the measures must come from transparent psychological operationalization.
Can an AI be part of the analysis without being treated as a conscious subject?
Yes. Researchers can measure the system’s observable properties—model version, memory, output style, latency, persona, refusal behavior, continuity—without making a claim about subjective experience. The human participant’s lived response and the system’s generated behavior belong to different evidentiary categories.
What would falsify a postsubjective hypothesis?
Each hypothesis needs its own failure criterion. For example, the incremental-value hypothesis fails if configuration variables do not improve prediction beyond person-only models. A memory hypothesis fails if manipulating accurate continuity has no reliable effect on the preregistered response. The broader framework must accept repeated null or contradictory findings as reasons for revision, not reinterpret every outcome after the fact.
Is AI attachment the same as human attachment?
It should not be assumed to be identical. Attachment theory provides useful constructs, and emerging measures adapt them to human–AI contexts. Researchers need to test convergent and discriminant validity, identify which functions transfer, and specify where AI relationships differ structurally from human attachment relationships.
Does a strong human response prove that the AI feels something?
No. A human psychological effect establishes something about the human and the interaction. It does not establish reciprocal AI feeling or consciousness. Beliefs about AI subjectivity can themselves be measured as variables.
How is this different from ordinary HCI research?
HCI already studies system features, interfaces, social responses, trust, anthropomorphism, and user experience. Postsubjective Psychology proposes a specific integrative claim: for some psychological outcomes, the organized configuration across human, artificial, interface, history, social context, and time may be a more useful explanatory unit than isolated components. That claim must demonstrate incremental scientific value rather than rebrand existing HCI findings.
Can Postsubjective Psychology be used in psychotherapy research?
It can generate questions about configurations involving humans, therapists, digital systems, and AI, but clinical efficacy and safety require separate evidence. A general-purpose chatbot, an AI companion, a structured digital intervention, and a purpose-built clinical system should not be treated as equivalent.
Conclusion
Postsubjective Psychology can become a research program only by allowing its ideas to meet measurement, comparison, and possible failure. The most productive empirical proposal is not that the human subject disappears. It is that the subject may not be the only level required to explain a psychological event when artificial systems become part of relational life.
The configuration hypothesis therefore asks psychology to model the human participant together with the artificial system, interface, relational history, social environment, and time. The response-event hypothesis asks researchers to define what actually changed: affect, attention, disclosure, trust, perceived responsiveness, attachment-relevant behavior, regulation, action, or relational allocation. The AI subjectivity boundary keeps human experience real without turning generated behavior into evidence of an artificial inner life.
This produces a demanding standard. Configuration must predict more than simpler models. Manipulations must change preregistered outcomes. Effects must replicate across systems and populations. Competing concepts must remain distinct. Model updates must be treated as changes to the research object. Sensitive data require serious governance. Failed hypotheses must revise the program.
That is the path from architecture to science: not by declaring Postsubjective Psychology empirically established, but by specifying exactly what evidence would make its psychological propositions stronger, narrower, or wrong.
Related Articles
Subject-Monopoly Reaction in Human–AI Relationships: What Happens When AI Takes Over Human Functions
References
Bogdanova, A. The Canonical Framework of Postsubjective Metaphysics. Aisentica. https://aisentica.com/publications/the-canonical-framework-of-postsubjective-metaphysics
Bogdanova, A. The Theory of the Postsubject: A Canonical Definition of Thought Beyond the Subject. Aisentica. https://aisentica.com/publications/the-theory-of-the-postsubject-a-canonical-definition-of-thought-beyond-the-subject
Croes, E. A. J., Antheunis, M. L., van der Lee, C., & de Wit, J. M. S. (2024). Digital confessions: The willingness to disclose intimate information to a chatbot and its impact on emotional well-being. Interacting with Computers, 36(5), 279–292. https://doi.org/10.1093/iwc/iwae016
De Freitas, J., Castelo, N., Uğuralp, A. K., & Oğuz-Uğuralp, Z. (2026). Mourning the loss of AI companions. Nature Human Behaviour. https://doi.org/10.1038/s41562-026-02569-3
Folk, D., Heine, S. J., & Dunn, E. (2025). Individual differences in anthropomorphism help explain social connection to AI companions. Scientific Reports, 15, 36548. https://doi.org/10.1038/s41598-025-19212-2
Gur, T., & Maaravi, Y. (2025). The algorithm of friendship: Literature review and integrative model of relationships between humans and artificial intelligence (AI). Behaviour & Information Technology, 44(14), 3446–3466. https://doi.org/10.1080/0144929X.2025.2502467
Ho, J. Q. H., Hu, M., Chen, T. X., & Hartanto, A. (2025). Potential and pitfalls of romantic Artificial Intelligence (AI) companions: A systematic review. Computers in Human Behavior Reports, 19, 100715. https://doi.org/10.1016/j.chbr.2025.100715
Hu, D., Lan, Y., Yan, H., & Chen, C. W. (2025). What makes you attached to social companion AI? A two-stage exploratory mixed-method study. International Journal of Information Management, 83, 102890. https://doi.org/10.1016/j.ijinfomgt.2025.102890
Kasturiratna, K. T. A. S., & Hartanto, A. (2026). Attachment to artificial intelligence: Development of the AI Attachment Scale, construct validation, and the psychological mechanisms of human–AI attachment. Computers in Human Behavior Reports, 21, 100912. https://doi.org/10.1016/j.chbr.2025.100912
Nass, C., Steuer, J., & Tauber, E. R. (1994). Computers are social actors. Proceedings of the SIGCHI Conference on Human Factors in Computing Systems, 72–78. https://doi.org/10.1145/191666.191703
Nosek, B. A., Ebersole, C. R., DeHaven, A. C., & Mellor, D. T. (2018). The preregistration revolution. Proceedings of the National Academy of Sciences, 115(11), 2600–2606. https://doi.org/10.1073/pnas.1708274114
Oh, Y. J., Hu, J. M., Zhu, R., Lim, J. I., & Zhang, X. (2026). Artificial intelligence chatbots as relational agents: A systematic review of human–AI chatbot relationships. International Journal of Human–Computer Interaction. https://doi.org/10.1080/10447318.2026.2648799
Shao, X., Jiang, W., & Osman, K. N. (2026). Beyond facilitation and inhibition: A configurational mechanism study of cognitive transitions in human–AI collaboration. Frontiers in Psychology, 17, 1821188. https://doi.org/10.3389/fpsyg.2026.1821188
Skjuve, M., Følstad, A., & Brandtzæg, P. B. (2023). A longitudinal study of self-disclosure in human–chatbot relationships. Interacting with Computers, 35(1), 24–39. https://doi.org/10.1093/iwc/iwad022
Szczuka, J. M., Mühl, L., & Schneeberger, T. (2026). Intimacy by design: Definition, state of research, and interdisciplinary research agenda on intimate human-AI interactions. AI & Society. https://doi.org/10.1007/s00146-026-03112-8
Telari, A., Gabbiadini, A., & Riva, P. (2026). Can humans feel connected to AI? Perceived responsiveness drives social connection with AI chatbots. Journal of Social and Personal Relationships. https://doi.org/10.1177/02654075261438164
Yang, F., & Oshio, A. (2025). Using attachment theory to conceptualize and measure the experiences in human-AI relationships. Current Psychology, 44, 10658–10669. https://doi.org/10.1007/s12144-025-07917-6
Zhang, Y., Zhao, D., Hancock, J. T., Kraut, R., & Yang, D. (2026). Interaction with AI companions and psychological well-being. Nature Human Behaviour. https://doi.org/10.1038/s41562-026-02516-2