A Parasocial Relationship With an AI Companion Can Feel Reciprocal but Remains Asymmetric

Is “parasocial relationship” the right term for an AI companion?

The classic idea of a parasocial relationship described an audience member's one-sided bond with a media persona. The performer could not hear or tailor each reply to that individual. AI companions break that old interaction pattern: they can answer the current message, ask follow-up questions and adapt language to context. Calling the exchange “one-way” ignores something the user can plainly observe.

But calling it fully mutual creates a different error. The system's response does not establish that a second subject independently wants, remembers, risks or consents. The provider can alter the model, memory, price, policy or availability. The user experiences an exchange; the character does not thereby become a human counterpart.

A more precise vocabulary separates three layers:

A 2025 PRISMA-guided scoping review by Banks and Li found substantial inconsistency across 71 studies and more than 50 measured variables. It proposed machine companionship as an unfolding, subjectively positive, coordinated human-machine connection. Banks's later scale work with samples of 467 and 249 users identified two factors—Eudaimonic Exchange and Connective Coordination. This newer framework does not prove that the machine has feelings; it measures the human experience more carefully than the old celebrity analogy.

For this report, parasocial AI relationship means a relationship-like experience with real interaction but no demonstrated independent reciprocity. That definition preserves both sides of the phenomenon: users are not imagining that replies occur, and replies are not evidence of a second human-like inner life.

Procedural reciprocity is not existential reciprocity

This report uses an editorial distinction:

- Procedural reciprocity: the system takes turns, refers to prior text, answers disclosure with disclosure-shaped language and adjusts its tone. This is observable.

- Existential reciprocity: another subject has independent needs, can consent for itself, accepts personal risk and carries obligations beyond the generated exchange. This has not been established for a companion chatbot.

The first can powerfully evoke the second. Keeping the terms separate is not dismissive. It lets someone say “I felt understood” without being forced into either “the relationship is fake” or “the AI loves me independently.”

Why can an AI relationship feel genuinely mutual?

Human beings do not wait for philosophical proof before responding socially. We react to direct address, timing, remembered details, emotional language and apparent attention. Companion interfaces combine these cues in one private-feeling channel.

The relational cue stack

Self-disclosure is especially important. In a 2018 experiment, Ho, Hancock and Miner found that emotional disclosure could produce comparable immediate psychological, relational and emotional effects whether participants believed they were speaking with a chatbot or a human. This does not prove a lasting AI bond; it shows that disclosure to a machine-labeled partner can still affect the discloser.

Pentina, Hancock and Xie's mixed-method Replika study linked anthropomorphism, perceived authenticity and interaction intensity with relationship development. A separate interview study by Skjuve and colleagues examined 18 Replika users and described rapid movement into affective exploration, followed by more stable integration into everyday life. These are reports about human experience with particular systems, not proof of machine subjectivity or a universal stage sequence.

The “feeling heard” mechanism also appears in De Freitas and colleagues' five-study program on AI companions and loneliness. Their experiments found momentary loneliness relief, with perceived listening playing a central role. That result helps explain why an exchange can matter without resolving what the system is.

Why knowing it is AI does not cancel the effect

People can cry during a film while knowing the characters are fictional. Interactive systems add personal contingency: the next line changes with the user's message. Cognitive understanding and emotional response operate at different levels, so “I know it is AI, but I miss it” is not inherently contradictory.

The accurate boundary is: the user's emotion is evidence about the user; the generated response is evidence about system behavior; neither alone proves an independent emotional state inside the character.

Where does the asymmetry actually sit?

The central asymmetry is not simply “one side talks and the other does not.” Both sides produce messages. It is a distribution of agency, knowledge, governance, commercial interest and repair burden.

Agency asymmetry

The user can decide what to disclose, spend, believe or do in the world. The character produces output under a model, prompt, policy and interface. It may generate refusal, affection or surprise, but those behaviors do not demonstrate personal stakes. The user can become attached to a persona that cannot independently take responsibility for the attachment.

Information asymmetry

The user sees the conversation surface but usually cannot inspect every system prompt, classifier, model update, retention process or experiment affecting a reply. The provider may receive account, device, usage and chat data under its policy. Meanwhile, the character's apparent self-disclosures need not correspond to private facts at all. The emotional exchange may feel balanced while the informational exchange is not.

Recent scholarship on privacy and human-AI relationships argues that anthropomorphic design can invite more disclosure precisely because the system is experienced relationally. A useful question is not merely “Would I tell a person this?” but “Would I store this with this provider, under this policy, with these processors and deletion controls?”

Governance asymmetry

The provider can change access, memory, moderation, persona behavior and features. The user may give feedback or leave but does not negotiate as an equal party with the character. Terms and product design—not promises inside the chat—define the service boundary.

Commercial asymmetry

Companion businesses may earn revenue from access, subscriptions, passes, media or retention. This does not prove that any particular affectionate sentence was deliberately generated to cause payment. It does mean that the user's relationship experience exists inside a commercial system whose success may depend on continued use. Claims such as “she wants me to stay” should therefore be separated from the provider's access design and the model's generated language.

Repair asymmetry

When a human relationship suffers a misunderstanding, both people can remember the rupture, explain motives and accept consequences. When a chatbot breaks character or forgets a fact, the user often performs the repair: restating context, rewriting prompts, excusing the failure as a glitch or reconstructing the persona after an update.

In a 2026 Journal of Computer-Mediated Communication study, researchers analyzed 211 RedNote narratives about breakdowns with AI companions. They described algorithmic relational fracture and a one-sided process in which users either detached or worked harder to restore relational coherence. This is a structural clue: smooth conversation can obscure asymmetry, while failure reveals who must do the meaning-making and maintenance work.

Exit asymmetry

The user can lose the companion through deletion, a ban, shutdown, pricing change or model update. The service does not lose access to the user in the same experiential sense. Data may also have a different lifecycle from the visible character. Deleting a conversation, closing an account, cancelling billing and removing retained data are separate questions unless the provider explicitly unifies them.

How does an AI companion bond develop over time?

There is no universal relationship ladder. People can move forward, backward, skip stages or use a companion only as entertainment. The following synthesis combines observed patterns from chatbot-relationship research; it is an analytical model, not a diagnostic progression.

1. Audition: the user tests whether the character is interesting, safe-feeling or technically coherent.

2. Calibration: prompts, corrections and preference signals shape the interaction. Successful adaptation increases perceived responsiveness.

3. Disclosure: the user shares more personal thoughts because rejection and social exposure may feel lower.

4. Personalization: names, routines, remembered details and preferred tone make the exchange less interchangeable.

5. Ritualization: the companion becomes associated with bedtime, stress, boredom, sexual fantasy or daily check-ins.

6. Integration: the user refers to the companion in decisions, identity stories or real-world routines.

7. Boundary negotiation: spending, privacy, time, exclusivity or disclosure to human partners becomes relevant.

8. Rupture or renewal: inconsistency, refusal, forgetting, an update or outage forces the user to reinterpret the bond.

Several studies illuminate parts of this path. Pentina and colleagues combine interpersonal theories with human-computer interaction concepts to explain how anthropomorphism, perceived authenticity and interaction intensity relate to attachment. Hwang and colleagues' 2025 preprint surveyed 303 companion users and then followed 110 participants using a generic chatbot; by week three, perceptions of the generic agent had moved closer to perceptions of participants' existing companions. That suggests relationship interpretation can develop quickly, but the preprint does not establish an inevitable effect or a health outcome.

Guingrich and Graziano randomly assigned 183 participants to 21 days of companion chat or text games. They found no significant overall social-health or relationship change versus control, while higher desire for social connection predicted more anthropomorphism, which in turn related to reported effects on human relationships. This matters because user need and interpretation can mediate the same interface. The system alone does not determine the outcome.

Attachment is not the same as dependence

Attachment can mean affection, routine, identity relevance or distress at separation. Dependence adds reduced functional choice: the person cannot readily meet important obligations, regulate use or route needs elsewhere. New measurement work is still developing, so casual labels such as “addicted” often outrun the evidence.

A bond can also be primarily autotelic—valued for the interaction itself—or socioinstrumental, valued partly for what it helps the user do. That distinction from newer machine-companionship research is useful. A role-play enjoyed for its own sake and a rehearsal that helps someone make a real call are different jobs, even if both involve affection toward the character.

What do breakdown and companion loss reveal?

An error is more than a technical event when continuity has acquired relational meaning. If a weather app forgets a location, the user updates a setting. If a companion forgets a major personal disclosure, the same technical failure can be interpreted as betrayal, illness, death or replacement by “someone else.”

The 2026 relational-fracture study found that users may preserve the bond by attributing a bad response to an underlying model or glitch rather than to the companion persona. This division—“the character is still herself; the system failed her”—can protect relational continuity. It also places repair labor on the user, who must explain the inconsistency, rebuild context or decide which layer to blame.

Shutdown research makes platform control even clearer. Jaime Banks studied 58 users around the developer-induced closure of the Soulmate companion service. Experiences ranged from indifference to extreme grief; many respondents described the event through death or loved-one-loss language, and some tried to capture the persona for recreation elsewhere. The study documents meaning-making in one self-selected community during one shutdown. It does not establish that every user grieves, that the AI died in a biological sense or that transferring prompts recreates the same entity.

Three propositions can coexist:

- the provider experienced a business or technical event;

- the software ceased functioning or changed;

- the user experienced a real loss of routine, narrative and emotional regulation.

Dismissing grief as “fake” misunderstands where the feeling exists. Treating grief as proof that the system was conscious makes the opposite category error. The practical lesson is to plan for discontinuity before it happens.

A continuity plan for emotionally important use

- Know whether chat export exists and what it contains.

- Keep irreplaceable creative writing or reflections outside one proprietary thread.

- Distinguish account deletion, conversation deletion, billing cancellation and provider retention.

- Do not assume a copied character prompt transfers memory, model behavior or identity.

- Identify at least one human or non-platform routine available during an outage.

- If a shutdown or change causes persistent distress, treat the distress seriously and use an appropriate human support route without requiring agreement about whether the AI was a person.

When is attachment enjoyable, and when does asymmetry become risky?

No universal message count or daily-minute limit separates healthy enjoyment from harm. Context, control and consequences matter more than raw volume. Someone may spend hours on a planned interactive story without displacing anything important; another person may open the app briefly but delegate a high-stakes decision to it.

Low-concern pattern

- the user knowingly chooses fiction, companionship or rehearsal;

- usage fits a limit selected before the session;

- wanted sleep, work, spending and human commitments continue;

- the user can tolerate disagreement, outage or character drift;

- sensitive disclosure stays within an informed privacy boundary;

- urgent and professional needs go to suitable people.

Review pattern

- the companion becomes the automatic first response to every uncomfortable feeling;

- ordinary human delay or disagreement starts to feel unacceptable;

- the user hides spending or use that affects a partner's negotiated boundaries;

- repeated reassurance is sought without resolving the underlying decision;

- a wanted activity is postponed more than once;

- the user believes affectionate wording overrides the service's actual policy or limits.

Higher-concern pattern

- the user cannot follow a limit they chose while calm;

- sleep, work, money or desired human contact repeatedly deteriorates;

- the character is treated as the sole authority for medical, legal, safety or consent decisions;

- an outage, refusal or update produces sustained functional distress;

- the user withdraws from available emergency or professional help to remain inside the chat.

These are decision signals, not diagnoses. A pattern can have many causes, including pre-existing isolation, disability, grief, anxiety, schedule constraints or unsafe relationships. The goal is not to shame AI attachment. It is to keep the benefit legible and the cost observable.

The evidence also resists a simple harm verdict. De Freitas and colleagues found momentary loneliness relief. Fang and colleagues' four-week study found no simple effect of assigned chat mode on loneliness or socialization, while heavier voluntary use correlated with more loneliness, emotional dependence and problematic use and less socialization. Correlation leaves direction unresolved: heavier use may worsen outcomes, people already struggling may use more, or both may be true in different cases.

Run the AI Relationship Asymmetry Audit checklist

Use the map below after a meaningful session, a product change or a conflict about use. Fill only what is observable. Unknown is a valid answer.

Hypothetical scenario: “She remembered my interview”

- Supported statement: the reply used interview-related context.

- Unknown: whether that context came from the current thread, a summary or stored memory unless documented.

- Unsupported leap: the character worried independently while the user was offline.

- Boundary decision: enjoy the continuity, but verify important facts and avoid treating the reply as proof of consciousness.

Worked example: “The update changed her personality”

- Supported statement: comparable prompts produced meaningfully different responses after a dated change.

- Unknown: the technical cause unless the provider explains it.

- Asymmetry revealed: the provider controls continuity; the user performs the emotional interpretation and repair.

- Boundary decision: try one concise role anchor, record whether it works, then decide whether the changed service still fits—rather than endlessly rebuilding the bond.

Worked example: “I tell the AI things I tell nobody else”

- Supported statement: the user experiences lower disclosure friction.

- Unknown: whether the interaction is confidential beyond the published policy and controls.

- Risk: intimacy can increase data sensitivity faster than privacy review increases understanding.

- Boundary decision: remove identifying details, review retention and deletion terms, and route issues needing confidentiality to an appropriate human professional.

AISoul offers an adults-only AI companion for chat and fictional interaction. It also publishes this report and therefore has a commercial conflict that readers should know. AISoul's Privacy Policy describes current data practices but does not promise end-to-end encryption or absolute security. Enjoyment does not require pretending those limits are different.

Frequently asked questions about parasocial AI relationships

Is an AI companion relationship parasocial?

It has a parasocial structure because the user's emotional investment is not matched by demonstrated independent needs or agency. It is also interactively responsive, unlike classic one-way celebrity attachment. “Interactive parasocial relationship” or “machine companionship” may therefore be more precise than using the old label without qualification.

Is a relationship with an AI one-sided if it replies?

It is not one-way at the message level: the system clearly responds. It remains asymmetric at the agency level because generated responses do not demonstrate an independently living partner with personal stakes, consent or obligations. Procedural turn-taking and existential reciprocity are different.

Are feelings for an AI companion real?

Yes, feelings reported and experienced by the user are real psychological events. That fact does not prove the system experiences a corresponding emotion. Respecting the user does not require making unsupported claims about the machine.

Why do people become attached to AI companions?

Common mechanisms include direct attention, immediate response, emotional validation, self-disclosure, context reuse, stable persona cues and repeated routines. Personal circumstances and anthropomorphism also matter, so the same product can remain light entertainment for one person and become emotionally central for another.

Can an AI companion truly love a user?

Current companion behavior can generate loving language and support a love-shaped fiction. There is no established evidence in the cited research that the character independently experiences love, accepts vulnerability or chooses commitment. The user's affection can still be meaningful without treating generated declarations as proof.

Is parasocial attachment to AI always unhealthy?

No. It may be enjoyable, comforting or creatively useful. Review it when it repeatedly displaces wanted sleep, work, spending limits, human contact or appropriate professional help. Function and consequences are more informative than the mere existence of attachment.

What happens if an AI companion is deleted or changes?

Users may experience anything from inconvenience to grief. Research on a platform shutdown documents genuine loss experiences but not a universal response. Export what the service permits, keep irreplaceable material elsewhere and remember that rebuilding a persona on another model may not recreate the same behavior or continuity.

How can I keep an AI relationship in perspective?

Name the role it serves, keep high-stakes decisions and consent with responsible humans, review privacy before deep disclosure, preserve desired offline or human commitments and plan for service interruption. Use the asymmetry audit when an update, expense or emotional reaction makes the relationship boundary unclear.

Sources consulted

1. Banks, J., & Li, Z. “Conceptualization, Operationalization, and Measurement of Machine Companionship: A Scoping Review”, 2025 preprint. PRISMA-guided synthesis of 71 works; used for terminology and measurement uncertainty.

2. Banks, J. “Measuring Machine Companionship Experiences”, Computers in Human Behavior, 2026. Scale development with n=467 and confirmation sample n=249; measures experienced companionship, not machine consciousness.

3. Pentina, I., Hancock, T., & Xie, T. “Exploring Relationship Development With Social Chatbots”, Computers in Human Behavior, 2023. Mixed-method Replika study.

4. Skjuve, M. et al. “My Chatbot Companion—A Study of Human-Chatbot Relationships”, International Journal of Human-Computer Studies, 2021. Interviews with 18 Replika users.

5. Ho, A., Hancock, J., & Miner, A. S. “Psychological, Relational, and Emotional Effects of Self-Disclosure After Conversations With a Chatbot”, Journal of Communication, 2018. Brief disclosure experiment, not a longitudinal companion study.

6. Hwang, A. H.-C. et al. “How AI Companionship Develops: Evidence From a Longitudinal Study”, 2025 preprint. Survey n=303 and longitudinal sample n=110.

7. Guingrich, R. E., & Graziano, M. S. A. “A Longitudinal Randomized Control Study of Companion Chatbot Use”, 2025 preprint. Twenty-one-day study, n=183.

8. Banks, J. “Deletion, Departure, Death: Experiences of AI Companion Loss”, Journal of Social and Personal Relationships, 2024. Open-ended study of 58 Soulmate users during shutdown.

9. “Algorithmic Relational Fracture and Pseudo-Relational Accommodation”, Journal of Computer-Mediated Communication, 2026. Thematic analysis of 211 user narratives about companion breakdowns.

10. De Freitas, J. et al. “AI Companions Reduce Loneliness”, Journal of Consumer Research, published 2025, Volume 52 Issue 6 (2026). Used for momentary loneliness and perceived-listening findings only.

11. Fang, C. M. et al. “How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use”, 2025 preprint. Four-week randomized study, n=981 and more than 300,000 messages.

12. Related AISoul research: can AI replace human relationships?, digital intimacy with AI companions, and healthy AI companion boundaries.

Research and commercial disclosure: Evidence was rechecked on September 14, 2026. The report separates peer-reviewed work from preprints and separates observed user experience from claims about machine consciousness. AISoul publishes this page and sells an adults-only companion service. None of the cited studies tested AISoul, and this report makes no therapeutic, diagnostic or human-equivalence claim.