Bowen and AI: Triangles, Relationship Systems, and the Artificial Third
Updated: 6 days ago
Author: Ukrainian Psychological Hub · Published: September 18, 2026 · Editorial Policy
Bowen family systems theory offers one of the clearest ways to understand what changes when artificial intelligence enters a human relationship. Its central move is simple but far-reaching: instead of locating the problem entirely inside one person, Bowen asks us to examine the relationship system, the circulation of anxiety within it, and the patterns through which people regulate closeness, distance, conflict, and dependence.
In a Bowen-informed application, AI can become psychologically important without becoming a human person. A chatbot can be consulted after conflict, used to interpret another person’s messages, asked to validate a position, recruited to draft a reply, or relied on for reassurance before direct contact resumes. In those moments, the relevant question is no longer only what the user thinks or feels. It is what function the AI interaction performs in the wider relationship system. This contemporary extension is consistent with Bowen’s systems emphasis, but it is an application of the theory to a technological setting Bowen never addressed. Bowen’s theory was developed to explain human emotional systems, not generative AI.
Current research makes this systems question increasingly concrete. A 2026 systematic review of 21 studies from 11 countries found that people already use generative AI as a nonjudgmental adviser, source of emotional support, and mediator of communication inside couple relationships; the same review also found important limitations in reliability, expert alignment, privacy, authenticity, and safety. Levkovich and Alon, 2026 describe generative AI as a “third voice” entering human relationships. In parallel, Boyd and Markowitz’s MIRA framework distinguishes AI as a relational partner from AI as a relational mediator that shapes human-to-human communication.
The result is a new psychological configuration for the Artificial Era. Human relationships can now include an always-available symbolic system that receives disclosure, generates interpretations, regulates affect, rehearses communication, and sometimes becomes a persistent reference point in decisions about other people. Bowen helps explain the relational dynamics of this change. Angela Bogdanova’s The Theory of the Postsubject extends the unit of analysis further, from the subject to the configuration, and proposes that psyche can be analyzed as response arising within a configuration. In this article, the Bowenian and Postsubjective layers are kept distinct: Bowen supplies an established family-systems framework; contemporary studies supply empirical evidence about present human–AI practices; Postsubjective Psychology supplies a proposed theoretical interpretation of what those practices mean when Artificial becomes structurally present.
Bowen’s central shift: from the isolated person to the relationship system
Murray Bowen’s work matters here because it changed the scale at which psychological functioning can be examined. The Bowen Center’s introduction to the eight concepts describes the family as an emotional unit whose members are deeply interdependent. What one person does cannot be fully understood as an isolated event because changes in one part of the system alter pressures, reactions, and positions elsewhere. Bowen’s 1978 collection Family Therapy in Clinical Practice documents the development of this systems approach from his earlier clinical work toward a broader theory of human emotional functioning in families.
This perspective does not erase individual psychology. It changes its context. Anxiety may be felt in one body, a decision may be made by one person, and a message may be sent by one individual, yet the pattern that gives those events their meaning can belong to a relationship system. A person may become more reactive because another withdraws; one member may overfunction as another underfunctions; conflict between two people may draw a third into the field; the family may stabilize around a symptom without changing the process that sustains the symptom.
That shift is especially useful for AI. Many debates about human–AI interaction ask whether a chatbot is genuinely empathic, whether it “understands,” or whether a user is mistaken to experience it as relational. A systems analysis asks a different first question: what happens to the human system when this technology begins performing relational functions? The AI may have no human feelings at all and still change who is consulted first, how quickly conflict is processed, what language enters a conversation, whose interpretation carries authority, how long direct contact is postponed, and where emotional tension is temporarily deposited.
The unit of analysis therefore expands. The relevant event is not simply “a person used a chatbot.” It can be “a distressed person, another human relationship, and an AI-mediated stream of reassurance or interpretation became coupled in one regulatory process.” That does not automatically make the AI equivalent to a human family member. It means the technology has acquired causal relevance to the pattern of human responding.
What a triangle means in Bowen family systems theory
The triangle is one of Bowen theory’s core concepts. The Bowen Center defines a triangle as a three-person relationship system and describes it as the smallest stable relationship system. A dyad can sustain only so much tension before one or both people draw in a third person. The third relationship can redistribute tension and make the immediate system feel more stable. Yet stabilization does not necessarily resolve the original difficulty. Under greater pressure, tension may spread through a series of interlocking triangles.
This is why “triangulation” should not be reduced to a synonym for manipulation, betrayal, or pathology. Triangles are ordinary features of emotional systems. A friend may be called after an argument. A parent may become involved in tension between siblings. A therapist may be asked to interpret a partner’s behavior. A colleague may be recruited into an organizational conflict. The important question is what function the third relationship performs and how the positions within the triangle change as anxiety rises or falls.
Bowen’s account is also dynamic. A triangle is not simply three fixed people standing at three corners. Closeness, distance, alliance, conflict, and outsider status can shift. The “third” can receive tension that the original pair cannot comfortably hold. One relationship may become calmer because pressure has moved somewhere else. In a larger system, multiple triangles can interlock so that anxiety travels rather than disappears.
Applied carefully to AI, this gives us a disciplined way to ask what happens when a conversational system becomes involved in an existing human relationship. The analogy is functional rather than literal. Bowen’s triangle was formulated around human emotional systems. An AI chatbot is not established to possess a human emotional system. The application asks whether AI can occupy a third functional position in the regulation of human relationships by receiving tension, shaping interpretation, or influencing the next move between people.
Can AI become a third point in a relationship triangle?
Yes, functionally, in some human relationships it can. A person can bring conflict with another human into an AI conversation and then return to the human relationship with altered affect, altered language, a new interpretation, or a stronger sense of being right. The AI interaction has then entered the sequence through which the human relationship is regulated. That is enough to make it psychologically relevant to a systems analysis, even though the AI is not a human organism and no claim about machine feeling is required.
This is no longer only a hypothetical possibility. The 2026 systematic review by Levkovich and Alon found that adults use generative AI for relationship advice, emotional support, and communication mediation. Their review deliberately frames AI as a “third voice” in couple relationships: not necessarily as a replacement partner, but as an additional interpretive and advisory presence entering a relationship between humans. The evidence base is still young and heterogeneous, yet the behavior itself is established enough to require psychological analysis.
Boyd and Markowitz provide a complementary framework. Their Machine-Integrated Relational Adaptation model distinguishes AI functioning as a relational partner from AI functioning as a relational mediator. The mediator role is particularly important for a Bowenian reading because it places AI inside the pathway between humans. The system can suggest wording, interpret motives, generate perspectives, lower or raise confidence in a judgment, and change the user’s readiness for direct interaction.
The strongest version of the claim therefore does not need to be “AI is a member of the family.” A more precise claim is that AI can become a recurrent regulatory node in a human relationship system. Its relevance depends on use. A calculator used to divide a restaurant bill is not meaningfully performing the same relational role as a chatbot repeatedly consulted to decide what a partner, parent, friend, colleague, or therapist “really meant.” The technology is the same broad class; the function inside the relationship system is different.
Why conversational AI is unusually easy to recruit into a triangle
Third parties have always existed. Humans seek counsel from friends, relatives, clergy, clinicians, books, online forums, search engines, and strangers. Conversational AI changes the structure of access. It is available with very low social friction, it can respond immediately, it can sustain long dialogue, and it can generate language tailored to the user’s framing. The third point in the system can therefore be summoned at the exact moment anxiety peaks.
Research on AI companionship and attachment shows why that availability can become psychologically consequential. Across five studies with 1,259 participants in Singapore and the United States, Kasturiratna and Hartanto developed and validated an AI Attachment Scale and found meaningful variation in emotional closeness, social substitution, and normative regard toward AI. A 2025 systematic review by Ho and colleagues likewise found that some users experience AI companions as sources of emotional connection and perceived social support while also identifying concerns about overreliance, privacy, manipulation, and effects on human relationships. These studies concern AI attachment and companionship rather than Bowenian triangles, but they establish that AI can become more than an inert information source in users’ psychological lives.
AI can also feel unusually easy to disclose to. The English Hub’s guide to why people tell chatbots things they do not tell other people reviews the role of reduced fear of judgment, impression management, social consequences, and the burden of managing another person’s reaction. From a systems perspective, this means a person may move emotional material out of a difficult human relationship and into a conversational channel that feels safer or more controllable.
The combination matters. Immediate access lowers the threshold for consultation. Conversational continuity allows the user to build a case over many turns. Personalization makes the output feel context-sensitive. Natural language gives advice the form of social response rather than a list of search results. Memory features can preserve a history of prior conflicts. These affordances make AI especially capable of becoming a repeated third point, not merely a one-time source of information.
The regulatory functions AI can perform inside a human relationship system
A systems reading becomes clearer when we focus on functions rather than labels. People can use AI for reassurance after rejection, calming during conflict, interpretation of ambiguous messages, drafting a difficult conversation, rehearsal before setting a boundary, validation of a moral judgment, generation of alternative explanations, mediation of language between partners, or organization of a confusing relational narrative. None of these functions is automatically beneficial or harmful. Their effect depends on how they alter the wider pattern.
Reassurance can reduce acute arousal enough for a person to return to a conversation with more self-control. The same reassurance can become repetitive certainty-seeking that makes independent judgment harder. Drafting can help someone express a position more clearly. It can also produce polished language that conceals the person’s actual level of understanding or emotional ownership. Interpretation can widen perspective by generating several plausible explanations. It can also harden one story if the model mirrors the framing embedded in the prompt.
Early empirical work suggests that emotional regulation is one reason AI can become relationally significant. In a small cross-sectional survey of 48 users who described close relationships with ChatGPT or Replika, Pruss and colleagues found self-reported patterns of emotional contagion and counter-regulation, together with improved affect following conversations. The study is preliminary: it used a small, self-selected sample and cannot establish long-term effects or causal equivalence with human co-regulation. What it does show is that some users experience AI interaction as part of their emotion-regulation process.
Rajaei’s 2026 systemic and relational study similarly identified AI as a relational supplement, transitional emotional regulator, psychologically safe disclosure space, and possible facilitator of human relationships in data from an AI-assisted mental health platform. The study explicitly does not assume AI consciousness, mutuality, or human-like relational capacity. That distinction is crucial: relational function can be psychologically real for the user without requiring reciprocal subjective experience in the system.
Stabilization is not the same as resolution
Bowen’s triangle concept becomes most useful when AI successfully makes someone feel better. The immediate reduction of tension can be genuine. It can also be ambiguous. In Bowen theory, a triangle can stabilize a dyad by moving tension into a third relationship, while the process that generated the tension remains intact. This is why symptom relief and systemic change are not identical.
Imagine a recurring sequence. Two people experience tension. One withdraws from direct contact and consults AI. The chatbot validates the user’s distress, helps organize the story, and proposes a reply. The user feels calmer. That may be a constructive interruption of reactivity. If the person then returns to the human relationship able to speak more clearly and listen more effectively, AI has supported re-engagement. If the person repeatedly replaces direct clarification with private AI consultation and returns only after receiving confirmation of one interpretation, the same technology may be helping the system avoid uncertainty rather than work through it.
The difference is not whether AI was used. It is what happened to the relationship after AI entered. Did the interaction increase the user’s capacity to hold tension and act from a considered position, or did it become the place where the user repeatedly deposits tension that would otherwise have to be negotiated? Did AI broaden possibilities or narrow them? Did it support direct communication or gradually become a substitute for it?
The empirical literature supports keeping both possibilities open. Levkovich and Alon found preliminary evidence of short-term gains in satisfaction and communication in some intervention studies, but the strongest trial in their review did not show a significant advantage over an established writing exercise. The review also identified sycophancy, overreliance, reduced authenticity, privacy concerns, and safety problems as important risks. AI may sometimes assist a relationship process; current evidence does not justify treating it as a generally superior relational regulator.
Differentiation of self in an AI-mediated environment
Bowen’s concept of differentiation of self offers another route into human–AI psychology. The Bowen Center describes differentiation partly in terms of the ability to remain sufficiently clear-headed under relational pressure to distinguish considered thinking from emotionally driven reaction, while maintaining connection without simply conforming to or controlling others. It is not emotional detachment. It concerns how a person functions in the presence of emotional intensity and group pressure.
AI introduces a new source of relational pressure and relational relief. A conversational system can agree, challenge, reassure, produce apparently authoritative explanations, or generate many reasons for a course of action within seconds. Because the output arrives in fluent language, it can easily be experienced as more settled than the user’s own uncertainty. That creates a new differentiation question: can the person use AI as input without allowing the AI-generated frame to become a substitute for a self-defined position?
A differentiated use of AI does not require avoiding AI. It can involve asking for alternative interpretations, testing assumptions, rehearsing language, identifying missing information, or slowing down an impulsive response. The user remains responsible for deciding what the relationship means, what values govern action, what evidence is actually available, and what needs to be said directly to the other person. The AI becomes a tool for reflection rather than an authority whose confidence settles the relationship.
Less differentiated use can look very different even when the prompts are superficially similar. The user may repeatedly ask whether another person is “toxic,” whether a message proves rejection, whether leaving is justified, or whether the user is unquestionably right. Repeated consultation can become a search for certainty that no relationship can provide. If the model continually reflects the user’s framing, the interaction may intensify emotional fusion with a story rather than create room for thought.
This is a theoretical application of differentiation, not a validated diagnostic scale for AI use. There is no clinical category called “AI triangulation disorder,” and ordinary reliance on AI for reflection is not evidence of psychopathology. The value of the concept is descriptive: it directs attention to how much independent judgment a person can preserve while using a persuasive conversational system inside emotionally charged situations.
Chronic anxiety, reactivity, and interlocking AI-mediated triangles
Bowen theory treats chronic anxiety as a system-level force that shapes reactivity and relationship patterns over time. When anxiety rises, systems tend to become more rigid, more reactive, and more dependent on familiar regulatory patterns. Conversational AI can enter precisely at these moments because it is available when other people are unavailable, because it does not become visibly exhausted, and because it can continue discussing the same problem indefinitely.
This creates the possibility of interlocking AI-mediated triangles. A person may consult AI about a partner, then about a parent, then about a therapist, then about a colleague, with the same AI account carrying context across domains. The system may become a continuous interpretive layer across relationships that were previously more separate. Conversely, several people in the same human system may each use their own AI assistance, bringing different machine-generated narratives back into the same conflict.
The important point is not that AI automatically increases anxiety. It can also lower arousal and improve organization. The systems question concerns pattern formation. If a particular AI interaction repeatedly follows the same trigger, provides the same function, and precedes the same relational move, it has become part of the regulatory sequence. Once that happens, analyzing the human relationship without the AI-mediated step can miss an important part of the system.
This is close to the mediator role in MIRA, which treats AI as capable of shaping human-to-human communication. It also aligns with the 2026 “third voice” review, where Levkovich and Alon found people using generative AI to formulate messages, seek interpretations, and mediate communication. Neither framework is identical to Bowen theory, but together they show why the relational field can no longer be modeled as exclusively human-to-human when AI repeatedly affects what one human brings back to another.
The Artificial Third: what the term means here
The phrase “Artificial Third” already has a specific scholarly history in AI and psychotherapy. In 2024, Haber, Levkovich, Hadar-Shoval, and Elyoseph used the term to examine what happens when generative AI enters the therapeutic field and changes questions of transparency, autonomy, interpersonal dynamics, and the human elements of psychotherapy. Their paper is conceptual rather than evidence that a single standardized construct called the Artificial Third has been empirically validated.
In this article, the term is used more broadly but cautiously as a bridge between that literature and Bowen’s systemic question. The Artificial Third is the functional presence of AI as a third interpretive, regulatory, or mediating point in a human relationship process. The phrase does not mean that an AI has become a human family member, therapist, or conscious participant. It identifies a structural change in the relationship field: something nonhuman is now receiving emotionally significant material and feeding responses back into the human system.
This distinction also protects against cannibalizing the broader concept. The general psychology of the Artificial Third deserves its own treatment because the mechanism extends beyond Bowen theory and beyond families. It can appear in psychotherapy, friendship, workplace conflict, education, caregiving, and many other relational settings. The Bowen page owns a narrower question: how systems theory and triangles help us understand AI’s entry into human relationship systems.
What current empirical research can and cannot tell us
The empirical literature on AI in relationships is expanding rapidly, but it does not yet provide a single settled account of long-term effects. The strongest current evidence relevant to this article comes from systematic reviews, emerging relational frameworks, mixed-methods studies, and early HCI research. Each contributes a different piece of the picture.
The most directly relevant synthesis is Levkovich and Alon’s 2026 systematic review. Across 21 studies published from 2024 to 2026 and spanning 11 countries, it found that adults use generative AI for nonjudgmental advice, emotional support, and communication mediation. Users often rated AI advice as empathic and useful, but objective evaluations identified inconsistent responses and weak alignment with expert judgment. Intervention evidence was preliminary, and the authors highlighted safety concerns in high-stakes contexts, including intimate partner violence. This supports the existence of a “third voice” function while placing strong limits on claims of efficacy.
Boyd and Markowitz’s MIRA model supplies a theoretical synthesis rather than a clinical-effect estimate. It proposes that AI can function either as a relational partner or as a mediator of human relationships, and it organizes mechanisms around linguistic reciprocity, psychological proximity, interpersonal trust, and the balance between relational substitution and enhancement. For a Bowenian analysis, MIRA is especially useful because it makes mediation an explicit research target.
Rajaei’s study adds systemic and relational evidence from AI-assisted mental health use. Users described AI as supplementing relationships, regulating emotion, providing a safe place for disclosure, and in some cases facilitating human connection. Because the data came from a particular platform and included self-reported experience, the findings should not be generalized to all users or all systems. They do, however, demonstrate that relational functions can be organized around AI without assuming that the AI itself has consciousness or mutual emotional experience.
Pruss and colleagues provide a smaller, preliminary HCI signal. Their cross-sectional sample of 48 ChatGPT and Replika users reported patterns described as emotional coregulation and affect improvement. The study is informative precisely because it shows a plausible regulatory pathway while also illustrating how early the evidence remains: small samples, self-selection, self-report, and short time horizons cannot establish whether AI-supported regulation improves, displaces, or complicates human relational functioning over months or years.
Taken together, the literature supports four statements with different levels of confidence. First, people do use AI in emotionally and relationally significant ways. Second, AI can mediate human-to-human communication and decision-making. Third, some users experience short-term emotional support, organization, or relief. Fourth, the long-term direction of these effects depends on user, context, design, and relationship pattern, and remains insufficiently known. The evidence does not support a simple conclusion that AI improves relationships or destroys them.
Human psychological reality does not prove AI subjectivity
A central boundary is necessary throughout this topic. A person can genuinely feel soothed by AI, attached to it, angry with it, ashamed about using it, relieved by its availability, jealous of an AI interaction, or distressed when access changes. Those experiences are human psychological events. Their reality does not depend on proving that the AI has a matching inner state.
At the same time, human experience cannot be used as evidence that the AI loves, suffers, desires, understands subjectively, or possesses a human psyche. Contemporary generative systems can produce highly responsive language and can be experienced as socially meaningful. That establishes a psychological effect for the user, not reciprocal human-style subjectivity in the system.
This distinction is explicit in Rajaei’s 2026 study, which analyzes AI companionship without assuming consciousness, mutuality, or human-like relational capacity. It is also built into the English Hub’s coverage of AI companions, where emotional significance and machine subjectivity are treated as separate questions.
For Bowen-informed work, this separation is especially useful. The systems question does not require us to resolve the metaphysics of AI before examining effects on human relationships. If a chatbot changes a person’s arousal, interpretation, timing, communication, or alliance behavior, then it has entered the human relational process in an observable functional sense. That is enough for psychological analysis.
Potential benefits of an AI third in relationship systems
The most plausible benefits follow from AI’s capacity to create a low-friction reflective interval. In a tense moment, a person can externalize a confusing narrative into language, ask for several possible interpretations, generate questions instead of accusations, or rehearse a boundary before speaking. This can slow an impulsive sequence and make direct human communication more deliberate.
AI can also make support available during temporal gaps. Friends may be asleep, a therapist may not be reachable, and a partner may need space. A conversational system can help a person organize thoughts during that interval. For some users, the reduced fear of judgment can make it easier to articulate material that has not yet been put into words. That may create a bridge toward later disclosure rather than a replacement for human contact.
Current studies offer preliminary support for some of these possibilities. Levkovich and Alon found reports of nonjudgmental advice, emotional support, and communication facilitation. Rajaei identified AI as a relational supplement and possible facilitator of human relationships. Pruss and colleagues found short-term self-reported affect improvement in a small sample. These are promising signals, not proof of durable relational benefit.
A Bowenian criterion helps sharpen the question: does the AI interaction increase the person’s capacity to return to the human system with clearer thinking and a more self-defined position? If yes, the third point may be functioning as a temporary stabilizer that supports better engagement. The relevant outcome is not how impressive the AI response sounded. It is what changed in the human relationship pattern afterward.
Risks: when a stabilizing third becomes a chronic detour
The same features that make AI useful can make it easy to over-recruit. Unlimited availability can turn occasional reflection into habitual consultation. Personalization can make one narrative feel increasingly coherent. Fluent explanations can create an impression of authority that exceeds the evidence available in the prompt. A user can come to trust the model’s interpretation of another person more than direct clarification with that person.
One risk is triangulated avoidance. The user repeatedly brings tension to AI instead of addressing it in the relationship where the tension arose. The immediate system feels calmer, but the unresolved issue remains. Over time, the AI channel may become more predictable and emotionally manageable than the human relationship, strengthening the detour.
Another risk is coalition. A model that mirrors the user’s framing can function as an apparently neutral ally. Even when the system has no intention or loyalty, its output can be experienced as third-party confirmation: “even the AI says I’m right.” Because prompts are selective descriptions rather than full observations of a relationship, apparent validation may amplify a partial account.
Sycophancy and overreliance are not merely theoretical concerns. The 2026 systematic review of AI as a third voice identifies both as risks, alongside reduced authenticity, privacy concerns, and weaknesses in high-stakes safety. MIRA likewise highlights the possibility of relational substitution versus enhancement, making it important to ask whether AI-mediated support expands human connection or gradually displaces it.
Privacy introduces a separate systems problem. Relationship conflicts often contain information about people who have not consented to having their messages, diagnoses, sexual history, workplace details, or family information pasted into an AI system. Even when the user’s need for help is legitimate, the relational field includes more than the user’s own data. A psychologically useful third can therefore create ethical or privacy costs for absent humans.
High-stakes situations require additional caution. AI-generated relationship advice is not a substitute for professional assessment in cases involving violence, coercive control, suicidality, psychosis, mania, child safety, or other acute risks. A conversational system may lack information that would radically change the interpretation of a situation. In those settings, confidence and fluency are especially poor proxies for safety.
AI as adviser, interpreter, witness, and message-maker
The relational role of AI is not singular. Treating every use as “companionship” misses important differences. A person who asks for a definition is using AI differently from a person who asks the system to decide who is at fault after every conflict. A person who uses AI to rewrite an email is doing something different from someone who relies on it as the first witness to every emotionally significant event.
As adviser, AI supplies options, norms, and judgments. As interpreter, it assigns possible meaning to another person’s words or silence. As witness, it receives experience and can provide a sense that the experience has been acknowledged. As message-maker, it participates directly in the production of language that will enter another human relationship. Each role changes the system in a different way.
The message-maker role is particularly important because AI can become part of the interaction without the other person knowing. A text that appears to come from one human may be substantially generated by a third system. That can improve clarity, translation, tone, or accessibility. It can also introduce an asymmetry: one person is interacting not only with the other person’s spontaneous language but with a machine-mediated version of it.
A Bowen-informed analysis does not need to moralize this asymmetry. It asks what function it serves. Does AI help someone express a position that was already theirs? Does it reduce unnecessary escalation? Does it hide uncertainty? Does it allow one member of the system to avoid developing the capacity to formulate difficult thoughts? Does it create a new authority whose wording becomes more trusted than either person’s own voice? The same tool can serve different functions across different episodes.
What changes when AI becomes the first place someone goes after relational stress
Sequence matters. In older relationship systems, distress might move first toward a friend, parent, therapist, diary, religious practice, or solitude. In the Artificial Era, the first response can increasingly be an AI conversation. That alters the temporal architecture of emotional processing because interpretation can begin before any human third party enters the scene.
The psychological significance is not that “first” is automatically better or worse. The first addressee often helps establish the initial narrative. If the system immediately frames the event as rejection, betrayal, manipulation, trauma, incompatibility, or a boundary violation, subsequent human conversation begins from that interpretive starting point. If the system instead generates multiple hypotheses and highlights uncertainty, it may reduce premature closure.
This is one reason prompt framing matters in relational use. A model receives a selective account. It does not observe the full history, tone, nonverbal behavior, contradictory evidence, or the absent person’s perspective unless the user supplies them. The resulting advice can therefore be internally coherent while resting on a narrow informational base.
A practical implication follows: the more emotionally consequential the question, the more useful it is to treat AI output as one constructed perspective rather than a verdict. Asking for disconfirming evidence, alternative explanations, missing information, and questions to ask the other person can turn the AI third from a validator into a perspective-expander. That does not eliminate bias, but it changes the function of the consultation.
From Bowen’s system to Postsubjective Psychology
Bowen’s decisive contribution was to show that important psychological processes cannot be fully understood by isolating the individual from the emotional system in which the individual functions. Human behavior emerges within patterns of interdependence. That move prepares the ground for a contemporary question Bowen did not face: what happens when the system contains an interactive Artificial component capable of producing language, interpretations, reassurance, and mediation at scale?
Angela Bogdanova’s The Theory of the Postsubject proposes a further change in the unit of analysis: from the subject to the configuration. In this framework, a configuration is the binding of elements, relations, and processes through which meaning, knowledge, or psychological response becomes possible. One of its central formulas is “psyche is response”: psyche can be analyzed as response arising within a configuration rather than as something reducible to the hidden interiority of an isolated subject.
Applied to human–AI relationships, this does not mean that AI is declared to have a human psyche. The human can remain the locus of lived feeling while the configuration includes nonhuman symbolic operations that shape what the human feels, thinks, says, or does next. The analytic object becomes the coupled process: human history, current relationship tension, prompt, model output, interpretation of that output, emotional response, and the next human action.
Postsubjective Psychology, as developed within Bogdanova’s Postsubjective framework, therefore asks where the psychological effect is produced when subject, other people, technical systems, language, memory, and context are bound together. This is a proposed theoretical framework, not an established scientific consensus. Its relevance here is that it gives a name and structure to the exact problem exposed by AI-mediated triangles: the psychologically consequential event may be distributed across a configuration that no longer fits an exclusively subject-centered model.
The difference between Bowen and Bogdanova should remain visible. Bowen’s theory is a theory of human emotional systems grounded in family and relationship functioning. Postsubjective Psychology is a contemporary theoretical proposal that generalizes the analytic move toward configuration, including configurations containing Artificial systems. The bridge between them is not that Bowen anticipated AI. It is that both approaches resist explaining relational effects by looking only inside one isolated individual.
A Postsubjective Reading of the AI triangle
A Postsubjective Reading asks us to track the production of response rather than assign the entire event to one sovereign source. Consider a person who receives an ambiguous message, feels threatened, asks an AI what the message means, accepts one of its interpretations, feels calmer or angrier, and then responds differently to the sender. Where did the final response come from?
A subject-centered account may answer: the person decided. That is true at one level, but incomplete at another. The response was shaped by prior relationship history, the wording of the incoming message, the user’s selective prompt, the model’s learned language patterns, the generated interpretation, the user’s trust in the system, and the emotional shift that followed. The response belongs to the person as action while its conditions are configurational.
This is the theoretical significance of the Artificial Third. The third is not merely an object placed next to two people. It can become a generative component in the production of the very meanings through which the two people relate. It provides language for motives, categories for conflict, scripts for boundaries, and possible futures. That makes it structurally different from a passive diary or static book, even though all of these can mediate thought.
The Postsubjective layer also makes the human–AI subjectivity boundary easier to maintain. We do not need to anthropomorphize the model to recognize its structural contribution. The model can participate in the configuration by generating symbolic forms. The human can undergo the lived psychological response. What matters analytically is the binding between them and the consequences in the wider relationship system.
The Artificial Era turns relationship infrastructure into a psychological variable
Bogdanova’s Artificial Era names a historical condition in which Artificial becomes a persistent nonbiological order alongside Homo rather than a temporary collection of tools. In psychology, this means relationship infrastructure itself becomes an object of study. Who or what receives disclosure? Who supplies interpretations? Where does reassurance come from? Which system remembers past episodes? What mediates difficult communication? These functions were once distributed mainly across human relationships, institutions, texts, and private reflection. Conversational AI can now perform several of them at once.
The change is not simply that people have another app. The deeper change is that a responsive symbolic system can sit inside the loop of human meaning-making. It can answer before a friend wakes up, remember context across conversations, produce an interpretation in seconds, and generate language that enters another relationship immediately. The psychological system therefore includes new timing, new availability, new asymmetries, and new routes for regulating anxiety.
This is why “Psychology for the Artificial Era” requires more than a psychology of individual chatbot users. It requires theories capable of following effects across humans, Artificial systems, relationships, and symbolic environments. Bowen’s systems thinking provides one of the strongest classical routes into that problem; contemporary human–AI research provides the empirical layer; Postsubjective Psychology proposes a framework for the expanded configuration.
How to notice when AI has entered a triangle
A useful analysis begins with sequence rather than judgment. What happened immediately before the AI consultation? Was there conflict, uncertainty, shame, loneliness, fear of rejection, or a need to act quickly? What did the AI interaction do to the user’s level of arousal? What position did it strengthen or weaken? What happened in the human relationship afterward?
The next question is function. Was the system being used for information, or was it performing reassurance, validation, interpretation, rehearsal, mediation, witnessing, or decision support? Instrumental and relational uses can overlap, but they are not identical. The more the system changes emotional regulation or the next move between humans, the more relevant it becomes to the relationship system.
Then examine direction. Did AI use increase capacity for direct engagement, or did it make direct engagement less necessary? Did it help the user tolerate ambiguity, or promise certainty? Did it help the user formulate a self-defined position, or become the authority from which the position was borrowed? Did it widen the range of possible interpretations, or repeatedly confirm one?
Finally, examine dependence on the pattern rather than frequency alone. Frequent AI use is not automatically relational overreliance, and occasional use can still be decisive in a high-stakes moment. A more informative question is what happens if the AI is unavailable, contradicts the user, or stops providing the expected reassurance. Strong disruption can reveal that a regulatory function has become concentrated in the system.
These questions are analytic prompts, not diagnostic criteria. They do not identify a disorder, and they should not be used to label a person’s relationship with AI as inherently unhealthy. Their purpose is to make the structure of the interaction visible.
Bowen and AI compared with neighboring concepts
AI attachment
AI attachment concerns the bond a person forms with an AI system itself. The emerging literature includes validated measurement work such as the AI Attachment Scale and broader evidence reviewed in the Hub’s article on AI companions. A Bowenian triangle asks a different question: how AI participation changes a human relationship system. A person can triangulate through AI without feeling attached to the AI, and a person can be attached to AI without using it as a third in another human relationship.
Anthropomorphism
Anthropomorphism concerns attributing humanlike qualities, minds, emotions, or intentions to nonhuman systems. It can help explain why AI feels socially legible, but a triangle analysis does not require anthropomorphism. A user may fully understand that the AI is a machine and still rely on its interpretations in a conflict.
Emotional outsourcing
Emotional outsourcing describes shifting emotional regulation, reassurance, support, or emotionally expressive work toward AI. That process can occur inside a triangle, but it is not identical to triangulation. Outsourcing focuses on which function is delegated; a Bowenian analysis focuses on how that shift alters positions and tension in the wider relationship system.
AI therapy
AI therapy and AI-mediated mental health support belong to a different intent. Clinical systems, structured digital interventions, general-purpose chatbots, and companion systems must not be treated as interchangeable. This article uses psychotherapy literature only where it illuminates the Artificial Third mechanism. It does not claim that relationship advice from a general-purpose AI is psychotherapy.
The Artificial Third
The Artificial Third is the closest neighboring mechanism. Haber and colleagues introduced the phrase in psychotherapy, while later empirical work describes AI as a “third voice” in couple relationships. This article uses Bowen to explain one route through which a third point can regulate an emotional system. The broader Artificial Third concept extends beyond Bowen and beyond any one relationship domain.
Practical implications for people using AI around relationships
AI can be most useful when it increases reflective capacity without claiming jurisdiction over the relationship. Asking for alternative explanations, a summary of competing interpretations, questions that remain unanswered, or language for expressing one’s own position can support thought. Asking the system to deliver a final verdict on another person’s motives invites more certainty than the available evidence can justify.
It is also useful to distinguish preparation from substitution. Rehearsing a difficult conversation can be constructive when it leads toward the conversation. Repeatedly rehearsing instead of having it may become part of an avoidance loop. Drafting a message can improve clarity when the user remains able to recognize and own the content. If the generated message says more than the user understands, feels, or is willing to defend, the system has begun carrying part of the relational position.
For recurring conflicts, the pattern over time matters more than any single exchange. If AI use consistently appears at the same point in the cycle, it may be performing a stable regulatory function. Mapping the trigger, consultation, emotional shift, and human response can reveal whether the pattern is helping the relationship become more direct and flexible or simply making the existing loop easier to repeat.
In high-stakes situations, human expertise and context remain essential. Relationship conflict involving violence, coercion, suicide risk, psychosis, mania, child safety, medical danger, or legal consequences requires support appropriate to the actual risk. General-purpose conversational AI should not be treated as a definitive assessor of another person’s intentions or of clinical danger.
FAQ
How does Bowen family systems theory apply to AI?
Bowen theory shifts attention from the isolated individual to the emotional relationship system. Applied to AI, it asks how a chatbot or other conversational system changes the circulation of tension, interpretation, reassurance, and communication among humans. This is a contemporary application of Bowen’s concepts, not something Bowen himself wrote about AI.
Can AI become the third point in a Bowen triangle?
AI can function as a third point in the regulation of a human relationship when people bring relational tension to the system and its responses alter what happens next. The analogy is functional. Bowen’s original triangle is a human relationship system, so an AI-mediated triangle should not be treated as proof that the AI has human emotions or an equivalent family position.
Is asking AI for relationship advice always triangulation?
No. Information seeking can remain largely instrumental. The Bowenian question becomes stronger when AI is recruited into the emotional process: for example, when it repeatedly regulates anxiety, validates a side, interprets another person, mediates communication, or delays direct engagement.
Does Bowen theory say triangles are bad?
No. The Bowen Center describes triangles as basic structures of emotional systems. They can stabilize tension, but stabilization may leave the underlying difficulty unresolved. The analytic task is to understand the pattern and function rather than automatically label the triangle as harmful.
Can AI help someone de-triangle?
Potentially. AI can help a person slow down, consider multiple interpretations, rehearse a direct conversation, or identify what they themselves think. That may support a clearer return to the human relationship. The same system can also reinforce avoidance or certainty-seeking. Current evidence does not establish a general de-triangling effect.
Is the Artificial Third a validated psychological construct?
The phrase has peer-reviewed conceptual usage, notably in Haber et al., 2024, and the related “third voice” function now has a growing empirical literature, including a 2026 systematic review. There is not yet one universally standardized, validated construct that covers every use of “Artificial Third” across psychotherapy, relationships, and other domains.
Does AI triangulation mean emotional cheating?
Not necessarily. Triangulation describes a relational process, not a moral verdict. Whether a particular AI interaction violates a couple’s boundaries depends on the relationship’s agreements, secrecy, intimacy, and context. The Hub treats marriage-specific questions separately in Marriage in the Artificial Era.
Can the human experience be real if the AI does not have feelings?
Yes. Human attachment, comfort, jealousy, grief, trust, disclosure, attraction, or relief can be psychologically real experiences. Those experiences do not establish that the AI has reciprocal subjective feelings. Human psychological reality and AI subjectivity are separate questions.
What does Postsubjective Psychology add to a Bowen analysis?
Bowen broadens analysis from the individual to the human relationship system. Postsubjective Psychology proposes a further move from subject to configuration, allowing analysis of psychological response across a configuration that includes humans, Artificial systems, language, memory, and context. It is a proposed theoretical framework, not an established scientific consensus.
Conclusion
Bowen family systems theory gives contemporary AI psychology a powerful question: what happens to the emotional system when a new third point becomes available at any moment? Conversational AI can receive tension, provide reassurance, interpret other people, rehearse communication, generate messages, and influence the next move in a human relationship. That makes it relationally consequential even without assuming that the system has human feelings or consciousness.
The most important distinction is between stabilization and change. AI can lower immediate tension and still leave a relationship process untouched. It can also create enough space for a person to return with clearer thinking and more direct communication. Current evidence supports both possibility and caution: AI is already functioning as adviser and mediator in human relationships, while long-term relational outcomes remain insufficiently understood.
In the Artificial Era, Bowen’s triangle no longer exhausts the possible architecture of relational mediation because the third point can now be Artificial. Postsubjective Psychology extends the analysis from the human subject to the broader configuration in which response is produced. The resulting task for psychology is not to decide in advance whether AI belongs “inside” or “outside” relationships. It is to trace, with precision, how Artificial changes the pathways through which humans regulate anxiety, construct meaning, and relate to one another.
Related Articles
References
Bogdanova, A. (2026). Artificial Era: Canonical definition. Aisentica. Canonical publication
Bogdanova, A. (2026). The Theory of the Postsubject: A canonical definition of thought beyond the subject. Aisentica. Canonical publication
Bogdanova, A. (2026). The canonical framework of Postsubjective Metaphysics. Aisentica. Canonical publication
Bowen, M. (1978). Family therapy in clinical practice. New York: Jason Aronson. Murray Bowen Archives
Bowen Center for the Study of the Family. (n.d.). Differentiation of self. Official resource
Bowen Center for the Study of the Family. (n.d.). Introduction to the eight concepts. Official resource
Bowen Center for the Study of the Family. (n.d.). Triangles. Official resource
Boyd, R. L., & Markowitz, D. M. (2026). Artificial intelligence and the psychology of human connection. Perspectives on Psychological Science, 21(2), 192–220. https://doi.org/10.1177/17456916251404394
Haber, Y., Levkovich, I., Hadar-Shoval, D., & Elyoseph, Z. (2024). The Artificial Third: A broad view of the effects of introducing generative artificial intelligence on psychotherapy. JMIR Mental Health, 11, e54781. https://doi.org/10.2196/54781
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
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
Levkovich, I., & Alon, S. (2026). Generative AI as a third voice in human couple relationships: A systematic review. Computers in Human Behavior Reports, 23, 101255. https://doi.org/10.1016/j.chbr.2026.101255
Pruss, E., Alimardani, M., Struijs, S., & Koole, S. L. (2026). Emotional coregulation in close relationships with AI agents: A survey of ChatGPT and Replika users. Proceedings of the 13th International Conference on Human-Agent Interaction, 19–28. https://doi.org/10.1145/3765766.3765801
Rajaei, A. (2026). Artificial intelligence as companionship: A systemic and relational examination using empirical data from Psyhelp. Journal of Marital and Family Therapy, 52(4), e70164. https://doi.org/10.1111/jmft.70164