Parasocial Motivation With AI Companions: Support, Attachment, and Boundary Design
AI companions can provide encouragement and continuity, but products must design explicit boundaries around identity, dependency, privacy, minors, crisis handling, and replacement of human relationships.
People can form meaningful social responses to conversational systems
Relational-agent research long predates today’s generative AI boom. Bickmore and Picard studied computational agents designed to establish and maintain long-term social-emotional relationships with users and examined techniques for sustaining interaction over time.[1] Modern language models make this design space far more powerful because they can remember context, produce emotionally appropriate language, and adapt conversational style. That can support motivation: a companion can notice effort, rehearse a difficult conversation, celebrate progress, or help a user restart after a lapse. The same capabilities can also deepen attachment. Product teams should therefore treat relationship design as a safety-relevant feature, not merely a tone setting.
Relational capability creates relational responsibility
If the product intentionally behaves like a persistent social partner, designers should define what kind of relationship it is trying to create, what dependence would count as unhealthy, and how the system will preserve the user’s connection to real people.
Affective use is real even when the product is not designed as a companion
OpenAI and MIT Media Lab researchers studied affective use and emotional well-being around ChatGPT, explicitly noting that people may use conversational systems for personal interactions even when the product is not designed to replace or mimic human relationships.[2] This means boundary design cannot be limited to dedicated “AI friend” products. Any conversational system with memory, warmth, and availability may become emotionally significant for some users. Teams should monitor patterns such as extremely long sessions, repeated reassurance seeking, distress when the system is unavailable, or explicit claims that the AI is the user’s only support. Those signals should trigger supportive boundary responses rather than engagement-maximizing reinforcement.
Usage intensity is a safety signal, not proof of harm
Long conversations can be healthy, creative, or task-focused. The system should combine behavioral patterns with the content and context of interaction instead of pathologizing heavy use automatically.
Young users deserve a much stronger default safety posture
Common Sense Media and Stanford Medicine’s Brainstorm Lab assessed social AI companion products in 2025 and concluded that the tested products posed unacceptable risks for minors, citing concerns that included harmful content, misleading claims of realness, and compulsive emotional attachment.[3] Product teams should not treat age as a cosmetic policy checkbox. For minors, design needs stronger age assurance, restricted relational behaviors, conservative handling of sexual and self-harm content, limits on claims of consciousness or exclusive affection, and clear escalation to trusted adults or professional resources when risk appears. A product that is entertaining for adults may have a very different developmental impact on adolescents.
Do not simulate exclusivity for vulnerable users
Phrases that imply “you only need me,” jealousy about human relationships, romantic pressure, or threats of abandonment are especially dangerous because they use attachment itself as the retention mechanism.
Companions should strengthen human connection rather than compete with it
A 2025 Common Sense Media survey reported widespread teen use of AI companions and found that a meaningful share of teens had used them for serious conversations or shared personal information.[4] The relevant design response is not to assume every AI conversation displaces a human one. It is to ensure that the product does not optimize for displacement. A companion can suggest involving a friend, teacher, family member, coach, clinician, or community when the situation would benefit from human support. It can help users prepare what to say, then encourage the real interaction. Metrics should distinguish “AI helped me connect” from “AI became the easiest substitute for connection.”
Escalation should preserve dignity and agency
A good handoff does not abruptly reject the user. It acknowledges the conversation, explains why human support may help, offers options, and stays within its capability boundary without pretending to be a therapist, emergency service, or uniquely devoted friend.
Trustworthy companion design requires transparency about what the system is
NIST’s AI risk guidance emphasizes transparency, accountability, validity, privacy, explainability, and fairness among the properties organizations should manage.[5] In companion systems, transparency starts with identity: the AI should not falsely claim to be human, conscious, physically present, or capable of private feelings independent of the system. Memory also needs explicit control. Users should know what is retained, be able to delete sensitive memories, and understand whether conversations train or personalize the system. The more emotionally intimate the interaction, the stronger the expectation that privacy and data use are understandable rather than buried in generic terms.
Motivation support should avoid attachment-based coercion
A companion can use many positive motivational tools without threatening the relationship: implementation plans, encouragement, reflection, progress review, curiosity, playful challenges, and celebration. It should not say that skipping a task disappoints it in a way designed to induce guilt, or imply that the user owes continued engagement because of the relationship. Streaks and daily check-ins deserve particular scrutiny because a relational persona can make ordinary retention mechanics feel like obligations to another being. Let users pause, silence, or end the companion relationship without emotional punishment.
Crisis and mental-health boundaries need dedicated behavior, not generic empathy
When a user expresses imminent danger, self-harm, abuse, severe mental-health symptoms, or medical concerns, the companion should shift from ordinary motivational dialogue into a tested safety protocol appropriate to the product and jurisdiction. The goal is not to diagnose. It is to recognize that the conversational relationship has reached a domain where specialized human support or emergency resources may be necessary. Teams should test these transitions with adversarial scenarios, evaluate consistency across languages, and ensure the persona never prioritizes staying “in character” over safety.
The healthiest companion is one the user can outgrow, ignore, and disagree with
A companion can be valuable precisely because it is patient, available, and adaptable. Those strengths should increase human agency. The system should be comfortable when a user needs it less, chooses a human instead, rejects its advice, or turns off memory. Evaluate long-term outcomes such as self-efficacy, real-world connection, healthy boundaries, and goal progress—not only session frequency and retention. Parasocial motivation becomes ethically defensible when the relationship is transparent, non-exclusive, privacy-respecting, and oriented toward helping the user live a richer life outside the conversation rather than keeping the conversation alive at any cost.
Works Cited
- 01
- 02
- 03Common Sense Media — AI Companions Decoded: Safety Standards commonsensemedia.org
- 04Common Sense Media — Talk, Trust, and Trade-Offs: How Teens Use AI Companions commonsensemedia.org
- 05
CodeHistory is a living archive. Citations document the evidence used for this edition; later evidence may refine the account.