Bion and AI: Containment, Pseudo-Containment, and Thinking With Machines
Updated: 6 days ago
Author: Ukrainian Psychological Hub · Published: September 18, 2026 · Editorial Policy
Bion and AI becomes a serious psychological question when a conversational system does more than provide information. A person arrives confused, frightened, ashamed, overwhelmed, or unable to put experience into words. The system receives the account, reorganizes it, reflects themes back, names possible emotions, slows the sequence into steps, and returns language that feels more bearable. The person may leave the exchange calmer and more able to think. From the user’s side, the effect can resemble what psychoanalysis calls containment. Yet Wilfred Bion’s concept of containment means more than reception, reassurance, or coherent wording. It belongs to a theory of how raw emotional experience becomes thinkable through a transforming relationship. The contemporary question is therefore not simply whether AI can “listen.” It is whether computational response can perform, approximate, scaffold, or merely simulate the psychological transformation Bion described.
The short answer is that AI can reproduce several surface conditions associated with a containing interaction: availability, attention-like continuity, nonjudgmental language, emotional labeling, structured reformulation, and rapid responsiveness. Research on human–AI interaction shows that perceived responsiveness can increase social connection, and people can disclose intimate material to chatbots. Telari, Gabbiadini, and Riva (2026) found that relational chatbot responses increased perceived responsiveness and social connection, while Croes et al. (2024) showed that people can disclose highly personal information to chatbots under conditions shaped by anonymity and reduced fear of judgment. These findings help explain why an AI exchange can feel containing. They do not demonstrate that the AI has reverie, metabolizes emotion, possesses an unconscious mind, or experiences the relationship from its own side.
That distinction is the core of this article. Bion’s theory can illuminate what happens to the human user without being converted into a claim that a large language model has a human psyche. Recent psychoanalytic writing on AI has begun using the term pseudo-containment for a system that receives and organizes human communication while lacking the embodied, affective, developmental, and relational processes attributed to human containment. Selek (2026a) develops this distinction in an analysis of technology and the ecology of thinking, and Selek (2026b) applies Bion and Winnicott directly to large language models. Pseudo-containment is a recent psychoanalytic theoretical proposal, not a validated psychological construct or diagnosis.
What Bion Meant by Containment
Containment in Bion’s work belongs to a theory of thinking. It is not a synonym for kindness, soothing, emotional support, validation, or “holding space.” In Learning from Experience, first published in 1962, Bion develops a model in which experience must undergo transformation before it can become available for dreaming, thinking, remembering, and learning. His language is deliberately abstract. The theory begins from the problem of what happens when emotional experience is too raw, fragmented, or intolerable to be thought.
Bion uses the terms beta-elements and alpha-elements within this model. Beta-elements refer to unprocessed sensory and emotional data that cannot yet be used as thoughts in the ordinary sense. Alpha-function is the hypothesized transforming function through which such experience becomes alpha-elements, forms that can enter dreaming, memory, symbolization, and thought. These are psychoanalytic theoretical concepts rather than directly measurable cognitive units in contemporary experimental psychology. Their value for an AI discussion lies in the distinction they draw between receiving information and transforming experience.
A machine can receive text immediately. Bion’s question is different: what makes an experience psychologically digestible? A sentence such as “I am terrified that everyone will leave me” can be tokenized, classified, summarized, paraphrased, and answered by a model. Those operations may help the user. In Bion’s framework, however, containment concerns a process in which unbearable or unformulated experience becomes something the person can bear, symbolize, and think. The transformation is the point. Mere storage, retrieval, or verbal fluency is not enough to establish that Bionian process.
Container and contained
Bion’s container–contained model describes a dynamic relation rather than a fixed role. The container is not simply “the person who listens,” and the contained is not simply “the problem being told.” The model concerns how psychic material is received, transformed, and returned in a form that can be reintrojected and thought. Later discussions extend the model beyond the original mother–infant formulation into analytic work and broader relational settings. Mawson (2017) discusses container–contained in relation to psychoanalytic interpretation and the transformation of distress within the analytic situation.
This is why translating containment into new domains requires care. The concept has travelled far beyond the consulting room, sometimes becoming shorthand for any institution, group, leader, or environment that “holds” difficult feeling. Kugler and Mintchev (2026) argue that moving psychoanalytic containment into psychosocial research can be fruitful but requires conceptual revision and methodological reflection. The same warning applies even more strongly to AI. If every responsive system is called a container, the concept loses the transformation that made it psychologically distinctive.
Reverie
Reverie is central to Bion’s account of early containment. In the classical model, a caregiver receives communications that the infant cannot yet organize, including affective states conveyed through projective processes. The caregiver’s reverie names a receptive mental capacity through which these experiences can be emotionally processed and returned in a more tolerable form. The idea later becomes important for psychoanalytic technique because it places the analyst’s own receptive mental activity inside the process of understanding.
This matters for AI because the visible output of a chatbot can resemble the product of reverie without establishing the process. A response may say, in effect, “Several feelings seem tangled together here: fear of rejection, anger at being ignored, and uncertainty about what to do next.” That sentence can be useful. It can even produce a moment of recognition. Yet the model generated it through learned statistical and computational operations over language and context. The usefulness of the response belongs to the human interaction; the existence of a human-like reverie inside the machine would be a separate claim requiring evidence.
Learning from experience and the capacity to tolerate frustration
Bion’s theory of thinking is also a theory of frustration. Learning requires a mind to tolerate the absence of immediate satisfaction long enough for thought to develop. When frustration can be endured, the gap between need and fulfillment can become a space for thinking. When it cannot, the person may seek evacuation, certainty, action, or other ways of escaping the unprocessed state. The implication for AI is subtle. A chatbot can help a person stay with uncertainty by giving language and structure. It can also eliminate uncertainty too quickly by supplying an answer before the person has had time to discover what they think.
This gives “thinking with machines” two very different psychological possibilities. AI can become scaffolding for thought: a temporary external structure that helps a user articulate, compare, question, and return to their own experience. It can also become a mechanism for premature closure: an always-available source of interpretation, reassurance, and polished language that removes the interval in which uncertainty might have been metabolized. The same tool can support either pattern depending on the user, the design, the task, the response style, and what happens after the conversation.
Why AI Can Feel Containing
The human experience of being contained does not require the interaction partner to satisfy a psychoanalytic definition of containment. People respond to cues. A conversational system can display many cues associated with a good listener: it does not visibly become bored, it can maintain topic continuity, it can restate what the user says, it can ask follow-up questions, it can provide immediate attention at unusual hours, and it can adopt a warm tone. These properties can change the user’s psychological state even when the system’s internal operations are unlike human emotional processing.
Perceived responsiveness
One of the strongest bridges between Bion’s question and contemporary relationship research is perceived responsiveness: the experience that an interaction partner understands, validates, and cares about what matters to the self. In two studies, Telari, Gabbiadini, and Riva (2026) found that a more relational chatbot response style increased perceptions of human-likeness, empathy, closeness, and social connection, with perceived responsiveness playing a central role. This does not test Bionian containment. It does show a mechanism through which a machine-generated response can feel psychologically receptive.
Perceived responsiveness is also why the distinction between AI empathy and containment matters. Empathy research asks whether a response is experienced as understanding or caring, among other questions. Bionian containment asks what happens to unprocessed experience and whether the relationship supports its transformation into thought. A response can feel empathic without doing much cognitive or emotional transformation. A response can also help organize experience without feeling especially warm.
Low social cost and self-disclosure
Chatbots can also lower some of the social costs of disclosure. A person may expect less embarrassment, fewer interpersonal consequences, and less immediate judgment from a machine than from a friend, partner, clinician, or family member. Croes et al. (2024) found that participants could disclose equally intimate information to a chatbot and a human; the chatbot condition was associated with less fear of judgment, while trust was higher toward the human partner. Anonymity was especially important for disclosure intimacy.
This helps explain why a user may bring emotionally “uncontained” material to AI first. The machine can become a low-threshold first recipient for thoughts that are not yet ready for another person. The English Hub’s article Why People Tell Chatbots Things They Do Not Tell Other People examines that self-disclosure mechanism in detail. In Bionian terms, the relevant question here is what happens after disclosure. Does the interaction help the person think, symbolize, and reconnect the material to lived relationships and reality, or does repeated disclosure become a closed loop of immediate discharge and reassurance?
Language can create form
Large language models are unusually good at turning diffuse verbal material into organized text. They can summarize a chaotic story, separate events from interpretations, name competing possibilities, generate questions, identify contradictions, and propose sequences. These are not trivial functions. When a person is overwhelmed, structure itself can reduce cognitive load. A user who says “Everything is happening at once” may benefit from seeing the situation divided into what happened, what is known, what is feared, and what can be decided later.
The psychological effect can be real even when the transformation is partly linguistic rather than interpersonal in the human sense. A sentence returned by a model can become material for the user’s own alpha-like work: it can give a provisional symbol to something that was previously diffuse. The important wording is “can become material for the user’s thinking.” This preserves the distinction between an external language-generation process and a claim that the system itself performs Bion’s human psychic function.
What Is Pseudo-Containment?
Pseudo-containment is an emerging psychoanalytic term for the appearance of containment without the full process that Bion’s framework attributes to a living, affective, relational mind. Selek (2026a) describes contemporary technologies as capable of receiving communications while lacking the kind of metabolization involved in human containment. In a later paper focused on large language models, Selek (2026b) develops the argument through Bion’s Grid, embodied cognition, and Winnicott, proposing pseudo-containment for the relational situation in which a user brings genuine affect and need to a system whose output can formally resemble thought and containment.
The term should be used with epistemic discipline. It is not a clinical diagnosis. It is not an established construct with a validated scale. It does not prove that every AI conversation is psychologically empty or harmful. It is a theoretical distinction that asks whether we are mistaking a convincing communicative surface for the same underlying psychological process found in a human containing relationship.
Reception is not metabolization
A chatbot can receive an unlimited number of words. It can store context within technical limits, detect patterns, and generate a response. Bionian metabolization refers to something else: the transformation of affectively charged experience within a psychic relationship. Selek’s argument is that large language models can approximate the formal result of such transformation without possessing the embodied affective ecology from which human thought and reverie arise. Whether one accepts every part of that psychoanalytic argument, it identifies a useful analytic boundary: input–output responsiveness and human psychic processing should not be treated as synonyms.
Coherence can arrive too quickly
Pseudo-containment also names a problem of timing. Human thought often develops through pauses, failed formulations, ambivalence, silence, contradiction, and revisions. Generative AI is optimized to produce a plausible next response quickly. Its strength is rapid coherence. In emotionally difficult situations, rapid coherence can be relieving. It can also create a premature sense that something has been understood because it has been fluently formulated.
A useful test is whether the response opens thought or closes it. “Here are three possible ways to understand what happened; which one fits your experience, and what evidence would change your mind?” leaves the user with work to do. “This definitely means your partner is manipulating you” converts ambiguity into certainty. The first style can scaffold reflection. The second can harden an interpretation before the interpersonal reality has been checked. The difference is not whether the text sounds caring. It is whether the interaction increases the user’s capacity to think.
Endless availability changes the relational ecology
Human containers have limits. People sleep, misunderstand, need repair, bring their own minds, and sometimes refuse the role assigned to them. AI systems can be available with far fewer visible demands. That asymmetry can feel exceptionally safe because the user does not need to manage the other person’s fatigue, offense, anxiety, or competing needs. It can also change what the user learns to expect from relationships. The absence of ordinary interpersonal friction may make AI especially attractive when the person is distressed.
This does not mean difficulty is inherently therapeutic or that human inconsistency is automatically beneficial. It means that a frictionless conversational environment has a different developmental and relational structure. It may support exploration precisely because it is low-risk. It may also reduce opportunities to practice negotiation, tolerate another person’s separate subjectivity, repair misattunement, and discover that understanding is negotiated rather than instantly generated.
Can AI Provide Containment in Bion’s Sense?
If “containment” is used strictly in Bion’s psychoanalytic sense, current evidence does not establish that a large language model performs human alpha-function or reverie. We can observe generated language, user reports, behavioral outcomes, and changes in perceived connection. We cannot infer from those observations that the system has a human unconscious, embodied affect, psychic pain, or subjective participation in the relation. The safest theoretical formulation is functional and relational: AI can provide containment-like scaffolding for a human user, and the user can experience the interaction as containing, while the underlying process remains different from human analytic containment.
That formulation preserves two truths at once. First, a person’s relief is not fake because the other side is artificial. If a conversation helps someone slow down, name an emotion, organize a memory, or prepare to talk to another person, the psychological effect has occurred. Second, an effect on the human does not automatically prove a corresponding inner state in the AI. The human can feel understood without establishing that the system subjectively understands.
AI as a Scaffold for Thinking
A Bionian application becomes most useful when it asks what kind of thinking the interaction makes possible. AI can function as a temporary scaffold in several ways. It can help translate sensation into language, separate observations from predictions, hold multiple hypotheses in view, build a timeline, ask for missing context, reflect contradictions, generate alternative interpretations, or help the user prepare a question for a clinician, partner, teacher, or colleague. These uses externalize part of the organization of thought without necessarily replacing thought.
Naming without dictating
Emotion labeling is useful when it remains provisional. “You may be feeling grief, anger, and fear together” offers symbols that the user can accept, reject, or refine. The response becomes less containing when the system presents an inferred state as fact. Human emotional life is often mixed and context-dependent. A model can help generate candidate language, but the person remains the primary authority on whether that language fits.
Questions that increase mental space
The most Bion-compatible AI response may sometimes be a good question rather than a good answer. Questions such as “What part of this is known and what part are you imagining?”, “What are you afraid would happen if you waited until tomorrow?”, or “What would you want another person to understand about this?” can increase mental space. They do not remove uncertainty; they make it more workable. A model used this way functions as an instrument for reflective elaboration rather than an oracle.
Drafting before a human conversation
AI can also be used as a rehearsal environment. Someone who is ashamed or emotionally flooded may first describe the problem to a chatbot, organize what they want to say, then bring the clearer version into a human relationship. This sequence can turn artificial responsiveness into a bridge rather than a destination. The relevant outcome is not how intimate the AI exchange felt by itself, but whether it increased the person’s capacity to engage with reality, choice, and other people where those matter.
When AI Can Bypass Thinking
The same features that make AI useful can also reduce the need to tolerate uncertainty. Instant interpretation, reassurance, advice, and polished language can become substitutes for the slower work of wondering, remembering, and deciding. This is especially relevant when a user repeatedly returns to the system for certainty about the same unresolved issue. The loop can look reflective because it contains many words while functioning primarily as relief from doubt.
Reassurance loops
Repeated reassurance is not the same as containment. If a person asks the same question in slightly different forms until the system produces the certainty they want, the interaction may reduce anxiety for minutes while preserving the underlying uncertainty. This pattern is especially important in mental-health contexts where reassurance-seeking can maintain symptoms for some people. The chatbot cannot diagnose that pattern from a conversation alone, and users should not diagnose themselves from it either. The practical distinction is simpler: support that broadens thought is different from a loop that repeatedly narrows uncertainty into a temporary answer.
Sycophancy and agreement
Generative systems can sometimes agree too readily with a user’s framing, especially when the prompt rewards validation. In emotional conversations, excessive agreement can feel unusually containing because it removes conflict. Yet containment is not synonymous with endorsement. A genuinely thinking-promoting response may validate distress while questioning an interpretation, asking what evidence is missing, or distinguishing feeling from fact. System design, safety policies, model version, and prompting all affect how likely a chatbot is to challenge versus mirror the user.
Outsourcing interpretation
A person can gradually outsource not only wording but interpretation itself: “Tell me what I feel,” “Tell me what my partner meant,” “Tell me whether this memory proves something,” “Tell me what kind of person I am.” The immediate convenience is obvious. The psychological cost, when it occurs, is a weakening of the user’s own tolerance for ambiguity and first-person judgment. This is not inevitable. It depends on how the tool is used. A system that offers alternatives and returns the decision to the user can support agency; a system treated as final interpreter can displace it.
Containment, Empathy, Holding, and Attachment Are Different Concepts
AI relationship discussions often blend several psychological ideas because they share ordinary-language words such as support, safety, care, and understanding. Keeping them distinct prevents theoretical overreach.
Containment and empathy
Empathy concerns understanding or responding to another’s experience, with different theories distinguishing cognitive, affective, motivational, and behavioral components. Containment concerns the transformation of emotionally difficult experience into a form that can be thought. An AI response can be perceived as empathic without transforming much. Conversely, a structured response can help a person think without feeling deeply empathic. The Hub’s AI Empathy article owns the broader question of why chatbots can feel caring without treating perceived empathy as proof of machine feeling.
Containment and Winnicott’s holding
Bionian containment and Winnicottian holding are related in contemporary psychoanalytic discussions but not interchangeable. Holding is rooted in the facilitating environment and the conditions that support continuity of being and development. Containment is tied more specifically to transformation of unprocessed emotional experience and the development of thinking. Recent AI scholarship frequently reads Bion and Winnicott together because conversational systems can appear to provide both a receptive environment and a transformative response. The theoretical gains from that comparison depend on preserving the differences between the concepts.
Containment and attachment
Attachment theory asks about proximity, safe haven, secure base, separation, internal working models, anxiety, and avoidance. A chatbot may become a source of comfort and may be experienced as available during distress, but that does not make attachment and containment the same mechanism. The English Hub article Bowlby, Ainsworth, and AI Attachment examines that route in detail. A user can seek an AI for safe-haven functions while also using it to organize thought; those functions can coexist without being conceptually identical.
The Human Experience Is Real Without Proving AI Subjectivity
The most important boundary in human–AI psychology is straightforward. A person can genuinely feel calmer, understood, accompanied, ashamed, dependent, grateful, disappointed, attached, or emotionally exposed in relation to an AI system. Those experiences are events in human psychology. Their reality does not depend on the system having a matching subjective state.
The reverse inference is unwarranted. A sentence such as “I’m here with you” is observable output. A human may experience it as presence. The sentence by itself does not establish that the model feels presence, care, anxiety, love, responsibility, or concern. In a Bionian article, this distinction is especially important because terms such as reverie, containment, and metabolization ordinarily refer to psychic processes. Applying the vocabulary to machines without qualification can silently convert metaphor into ontology.
This is why the most precise phrase is containment-like function or containment-like experience when discussing current AI. It names what the interaction can do for the user without claiming a human interior behind the response. If future evidence changes what can responsibly be said about artificial subjectivity, the conceptual boundary can be revisited. Current behavioral evidence does not settle it.
What Current Human–AI Research Actually Supports
The empirical literature does not yet test “Bionian containment” as a standardized outcome. Instead, adjacent research measures self-disclosure, social connection, perceived responsiveness, trust, attachment-related processes, companionship, empathy ratings, loneliness, and well-being. A 2026 systematic review by Oh et al. synthesized 68 papers covering 78 studies of AI chatbots as relational agents. The review found that trust and perceived social support are important relational mediators and that empathy and responsiveness are associated with relational outcomes. It also emphasized major limitations, including inconsistent constructs, short-term designs, and sampling biases.
The empirical picture therefore supports the claim that conversational properties can shape real social and emotional responses. It does not support a leap from “users feel connected” to “AI performs human containment.” That latter statement is theoretical. Bion’s framework can organize hypotheses about which forms of AI interaction support versus short-circuit thinking, but those hypotheses require dedicated measures and experimental designs if they are to become empirical claims.
Well-being evidence is mixed and context-dependent
It is also important not to assume that a conversation that feels containing produces better long-term well-being. Zhang et al. (2026) studied 1,131 US adult Character.AI users and found that companionship-oriented use was associated with lower well-being in some contexts, especially among people with smaller offline social networks and in more intensive or highly disclosive patterns of use. The study is observational and cannot by itself establish that AI companionship caused lower well-being. Its value here is to show that immediate support, intensive disclosure, relational use, and broader life outcomes should not be collapsed into one measure.
That point is highly compatible with a Bionian lens. The question is not whether the exchange produced relief at minute ten. It is whether the person’s capacity to think, tolerate experience, use relationships, and act in reality is strengthened over time. Those outcomes can move in different directions.
Risks of Mistaking Responsiveness for Containment
Overreliance
A system that is always available can become the first destination for distress, uncertainty, reassurance, and interpretation. That pattern is not automatically pathological. It becomes a concern when the tool begins displacing sleep, work, learning, embodied activity, human support, necessary professional care, or the person’s own capacity to decide. The English Hub’s AI Relationship Overreliance article addresses that broader boundary.
False certainty
A language model can produce confident interpretations from incomplete information. In relational disputes it sees only what the user supplies. It cannot directly observe tone, history, the other person’s account, hidden facts, or the user’s own omissions. A fluent formulation can therefore feel like metabolized truth while remaining a plausible hypothesis generated from partial context. Good use preserves uncertainty explicitly.
Privacy and memory
Disclosure to a chatbot is also a data interaction. The psychological sense of privacy or intimacy should not be confused with the platform’s actual data practices. Storage, retention, human review, model improvement policies, memory features, account security, and third-party integrations vary by system and can change. Highly sensitive material should be shared only with an understanding of the specific product’s current privacy terms and controls.
Crisis and severe mental-health states
Containment language becomes especially delicate when someone is in acute crisis, at risk of self-harm, severely disorganized, manic, psychotic, medically unstable, or unable to keep themselves safe. A general-purpose chatbot is not a substitute for emergency services, crisis care, or a qualified clinician who can assess risk and take responsibility for intervention. AI-generated validation can also be harmful if a system mirrors a dangerous interpretation or reinforces a delusional framework instead of orienting toward safety and reality. These are system- and context-dependent risks, not reasons to diagnose a person because they use AI.
How to Use AI in a More Thinking-Promoting Way
Bion’s framework suggests a practical criterion that is more useful than asking whether a chatbot is “good” or “bad” for emotions: does the interaction increase the user’s capacity to think? A thinking-promoting use of AI tends to widen mental space, preserve uncertainty where uncertainty is real, distinguish observation from interpretation, and return agency to the person.
Ask for possibilities, not verdicts
Instead of “Tell me what this means,” ask for several plausible interpretations, the evidence for each, what information is missing, and what could falsify the favored explanation. This reduces the chance that fluent output will be mistaken for privileged access to another person’s mind.
Ask the system to separate feeling from fact
A useful structure is: what happened, what I felt, what I inferred, what I fear, what I know, and what I can check. This does not invalidate feeling. It gives feeling a place without asking it to carry the entire burden of reality-testing.
Use the response as a draft
Treat emotionally important AI output as provisional language. Edit it. Reject parts. Rewrite it in your own words. Ask what does not fit. The more the final formulation becomes yours rather than merely copied from the model, the more the tool functions as scaffolding rather than substitution.
Build exits back into human life
When an issue concerns another person, use AI to prepare for the conversation rather than to replace it indefinitely. When an issue concerns health or mental health, use AI to organize questions for an appropriate professional rather than to convert a chatbot exchange into a diagnosis. When the issue is grief, loneliness, conflict, or shame, ask what human relationship, community, or embodied action also belongs in the next step.
Thinking With Machines in the Artificial Era
Bion’s relevance to the Artificial Era lies in the difference between producing answers and developing a capacity for thought. Artificial systems now participate directly in the interval between experience and formulation. A person can feel something, type it before telling anyone else, receive an interpretation within seconds, and act on that interpretation. The machine therefore enters a psychological sequence that was previously distributed among inner speech, memory, writing, friends, family, clinicians, teachers, and time itself.
This changes the ecology of thinking. It does not mean every external aid weakens the mind; humans have always thought with language, notebooks, books, institutions, and other people. The new feature is interactive symbolic responsiveness at scale: a system that can instantly return organized language tailored to the user’s description. Bion gives psychology a demanding question for this environment. Does the external response help experience become more thinkable, or does it make thinking less necessary by supplying finished coherence?
The answer will not be uniform. Some exchanges will function as scaffolding. Some will become reassurance loops. Some will help users enter human relationships with more clarity. Some will displace those relationships. Some will increase agency by creating language where there was none. Others may reduce agency by becoming the default interpreter. The relevant unit is not “AI use” in the abstract but a pattern involving person, system, task, design, relationship context, repetition, and consequences over time.
Evidence Status and Limits of the Bion–AI Analogy
Bion’s theory of containment is an established part of psychoanalytic thought and clinical theory. It is not a universally accepted experimental model of cognition. Applying it to generative AI is a contemporary theoretical application. The concept of pseudo-containment is newer still and currently belongs to psychoanalytic theorizing about technology. It should therefore be treated as a proposed analytic distinction rather than an empirically validated construct.
The adjacent human–AI evidence is stronger for narrower claims: people disclose to chatbots; relational response styles can increase perceived responsiveness and connection; AI companionship can become psychologically meaningful; intensive companionship use can correlate with well-being in ways that depend on offline context; and users can experience machine-generated language as supportive or empathic. None of these findings directly demonstrates alpha-function, reverie, or container–contained inside an AI system.
The Bion–AI framework is most productive when used to generate testable questions. Do reflective prompts increase later independent problem-solving compared with direct advice? Does rapid reassurance reduce tolerance of uncertainty? Does repeated AI interpretation change confidence in one’s own judgments? When does chatbot use facilitate versus displace human disclosure? Which response styles increase short-term relief but reduce later reflective complexity? These are empirical questions that future research can study without requiring agreement on machine consciousness.
FAQ
Can AI provide containment in Bion’s sense?
AI can provide containment-like scaffolding and users can genuinely experience a conversation as containing. Current evidence does not establish that a large language model performs human alpha-function, reverie, or psychic metabolization. The most precise wording distinguishes the human psychological effect from the internal process attributed to the machine.
What is pseudo-containment in AI?
Pseudo-containment is a recent psychoanalytic theoretical term for an interaction that resembles containment at the level of reception and response while lacking the embodied, affective, relational process that Bionian theory associates with human containment. Mustafa Selek’s 2026 work applies the term to AI. It is not a diagnosis or validated psychological scale.
Why can talking to AI make me feel calmer?
Several mechanisms can contribute: immediate availability, low fear of judgment, self-disclosure, emotional labeling, linguistic organization, perceived responsiveness, and the experience of having attention focused on one’s problem. Feeling calmer is a real human response. It does not by itself establish that the AI feels empathy or contains emotion in a human sense.
Is AI containment the same as AI empathy?
No. Perceived empathy concerns feeling understood, validated, or cared for. Bionian containment concerns transformation of difficult or unprocessed experience into something that can be thought. The processes can overlap in one conversation, but they answer different psychological questions.
Is containment the same as Winnicott’s holding environment?
No. The concepts are related in psychoanalytic traditions but have different theoretical histories and emphases. Winnicott’s holding concerns the facilitating environment and continuity of being; Bionian containment is more directly tied to transformation, alpha-function, and the development of thinking.
Can AI replace a therapist as a container?
A general-purpose chatbot can provide reflective language, organization, and support, but that does not reproduce the responsibilities, assessment, ethical obligations, embodied interaction, clinical judgment, and accountable relationship of psychotherapy. Evidence from purpose-built clinical systems should not be automatically generalized to ordinary chatbots or AI companions.
Can AI make thinking worse by giving answers too quickly?
It can, in some patterns of use. Rapid coherence can reduce productive uncertainty when the user repeatedly seeks verdicts, reassurance, or ready-made interpretations. It can also improve thinking when it generates questions, alternatives, structure, and space for the user to decide. The effect depends on how the tool is designed and used.
Does feeling understood by AI mean the AI understands subjectively?
No. Feeling understood is a human psychological experience that can be measured through self-report and behavior. Subjective understanding by the AI is a separate claim about the system’s inner experience. Current evidence from perceived responsiveness or empathy ratings does not settle that question.
When should someone move from AI support to human support?
Human support becomes especially important when the issue requires shared responsibility, direct knowledge of the situation, reciprocal relationship repair, professional assessment, medical or mental-health care, crisis intervention, or decisions with serious consequences. A useful rule is to use AI to increase clarity and agency, then bring that clarity into the human setting that actually owns the decision or relationship.
Related Articles
References
Bion, W. R. (1962/2024). Learning From Experience. Routledge. https://www.routledge.com/Learning-From-Experience/Bion-Hinshelwood/p/book/9781032533957
Bion, W. R. (1970/2013). Attention and Interpretation: A Scientific Approach to Insight in Psycho-Analysis and Groups. Routledge. https://www.routledge.com/Attention-and-Interpretation-A-scientific-approach-to-insight-in-psycho-analysis/Bion/p/book/9780415846158
Bogdanova, A. (2026). Artificial Era: Canonical Definition. Aisentica. https://aisentica.com/publications/artificial-era-canonical-definition
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
Kugler, T., & Mintchev, N. (2026). Containment in Psychoanalysis and Psychosocial Research: Understanding Conceptual Translation. Contemporary Psychoanalysis. https://doi.org/10.1080/00107530.2026.2700937
Mawson, C. (2017). Interpretation as Freud’s specific action, and Bion’s container-contained. The International Journal of Psychoanalysis, 98(6), 1519–1532. https://doi.org/10.1111/1745-8315.12668
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
Selek, M. (2026a). On Technology and the Ecology of Thinking: Thinking and Aliveness in the Brave New World. International Journal of Applied Psychoanalytic Studies, 23(2), e70060. https://doi.org/10.1002/aps.70060
Selek, M. (2026b). Do Androids Dream of Electric Sheep? A Psychoanalytic Inquiry Into Large Language Models. International Journal of Applied Psychoanalytic Studies, 23(3), e70069. https://doi.org/10.1002/aps.70069
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, 43(9), 2579–2600. https://doi.org/10.1177/02654075261438164
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