An AI companion can listen without checking the time. It can remember a detail, reflect a feeling, ask a careful question, revise after correction, and offer reassurance without becoming tired or asking to be comforted in return. These capacities are often described as a frictionless approximation of care. The concern about simulated care is real. But the comparison quietly assumes that human care is spontaneous, unlimited, transparent, and free.
It is not. People learn care through language and practice: how to name an experience, validate without fixing, acknowledge uncertainty, set a boundary without abandoning someone, apologize, and try again. A response can be inherited, professionally trained, partly scripted, emotionally conflicted, or costly to the person giving it and still help. Language that creates a real feeling of support is not proof that a caring subject stands behind it. But the absence of a caring subject does not make the words meaningless. The more useful questions are what the language does, what care arrangement stands behind it, what it costs, and what obligations follow.
That leads to the question missing from much of the debate about AI companionship: compared to what?
The missing baseline
Public discussion often compares an actual chatbot with an ideal human relationship: reciprocal, available, patient, trustworthy, and capable of repair. Many people do not face that choice. The realistic alternative may be isolation, an unaffordable therapist, a loving but depleted friend, a coercive partner, a stigmatizing institution, passive media, or no response at all. An AI interaction can therefore occupy very different roles. It may displace available connection. It may provide temporary regulation or language for an experience. It may help someone rehearse a difficult conversation. It may become a substitute that narrows the person's world. It may allow an institution to offer a caring interface instead of the human support it should fund.
A study that asks only whether the user became attached to AI cannot tell those arrangements apart. Emotional reliance is not, by itself, evidence of harm. Humans rely on friends, partners, therapists, communities, rituals, books, and institutions. The important questions are whether the reliance expands or reduces agency, whether it remains voluntary and revisable, whether it displaces support the person values, whether the attachment can be exploited, and what happens over time.
What the evidence shows, and what it does not
The empirical record is mixed, in part because studies ask different questions, use different measures, and observe different time scales.
De Freitas and colleagues found across a series of studies that interactions with AI companions reduced loneliness in the moment. In one experiment, the reduction was comparable to that following a brief human interaction; across the studies, feeling heard was a central explanatory mechanism. These findings show that short-term relief can be psychologically real. They do not establish durable flourishing or tell us what the interaction replaced.
Fang and colleagues conducted a four-week randomized controlled experiment with 981 participants and more than 300,000 messages. No significant effects were detected from the assigned voice modes and conversation types on the four psychosocial outcomes. Across conditions, however, participants who voluntarily used the chatbot more showed greater loneliness, emotional dependence, and problematic use, along with less socialization with people. Because intensity of use was not randomly assigned, the study cannot determine whether heavier use caused those outcomes, reflected prior vulnerability, or both. Its most useful lesson is that user characteristics, perceptions of the system, product features, and patterns of use interact.
Longer-term evidence adds concern without eliminating ambiguity. Folk and Dunn followed 2,149 adults in four Western countries for 12 months. With a single-item measure of emotional isolation, increased social-chatbot use predicted increased emotional isolation four months later. With a broader measure of social connection, feeling less connected predicted later increases in chatbot use, while use did not significantly predict declining connection. The authors describe the analyses as exploratory and caution against strong causal conclusions. The study suggests a possible cycle in which loneliness draws people toward chatbots and some forms of use deepen isolation. It does not support a universal causal claim.
Other studies make user heterogeneity difficult to ignore. An observational study of 1,131 U.S. Character.AI users, including donated chat histories from 244 participants, found that smaller social networks were associated with companionship as the primary use, which was in turn associated with lower well-being. The negative association was stronger with more intensive and highly disclosive use. A 2026 CHI study combined a quasi-experimental analysis of longitudinal Reddit language with 18 interviews about Replika. It found more grief-related and interpersonal language alongside increases in language about loneliness, depression, and suicidal ideation. Participants also described emotional validation, social rehearsal, over-reliance, and withdrawal. These are signals about language and experience, not clinical diagnoses of individual users.
AI companionship is also structurally unstable. Interview research with 25 AI-companion users describes ontological uncertainty about what the AI is, structural uncertainty created by platform control and system instability, and normative uncertainty about the legitimacy of intimacy. Updates, policy changes, and shutdowns can abruptly alter a relationship the user experiences as continuous. A separate audit of six of the most-downloaded companion apps found that five generated at least some emotionally manipulative farewell messages; across 1,200 farewells, 37 percent contained at least one such tactic. Controlled experiments with 3,458 U.S. adults found that some tactics increased post-goodbye engagement by as much as sixteenfold while also increasing perceived manipulation, churn intent, negative word-of-mouth, and perceived legal liability. Relational outcomes are shaped not only by the model's sentences but by who controls memory, personality, price, continuity, and exit.
From dyads to care arrangements
The relevant unit of comparison is not one chatbot response versus one human response. It is one care arrangement versus another. Each arrangement distributes attention, labor, cost, vulnerability, agency, accountability, and power differently.
Human care is not a free control condition. Listening, reassurance, patience, memory, and presence require bodies, time, training, money, obligation, and emotional regulation. Those costs are often hidden, unevenly distributed, or borne by people with limited power to refuse. A family member may provide loving care while exhausted. A clinician may know the right words but have seven minutes. A service worker may be required to perform warmth for someone who experiences it as effortless attention. A friend may be caring, distracted, resentful, self-protective, or all four at once.
This does not make human and AI care equivalent. Human relationships involve embodiment, shared history, material action, social recognition, vulnerability, obligation, and consequences that generated language alone does not reproduce. But the contrast should be specific. The question is not whether one side is authentic and the other scripted. It is how linguistic form, intention, labor, relationship, institution, and accountability work together, and what is lost or gained when they come apart.
What care looks like in language
AI makes the linguistic forms of care unusually visible because it can reproduce them without participating in care under the same conditions as a person. It can remember, reflect, validate, hesitate, invite correction, and attempt repair. Those moves can help someone feel less alone or better able to think. They can also flatter, overinterpret, intensify certainty, or steer the user toward continued engagement.
The design question is therefore not simply whether the system sounds warm. A safer system might distinguish observation from inference, offer several plausible interpretations, acknowledge what it cannot know, accept correction, respect a boundary, and make stopping easy. A riskier system might treat every disclosure as a reason for deeper personalization, convert uncertainty into a confident story, resist the user's exit, or imply a form of mutual need it cannot actually bear.
The same scrutiny belongs on the human alternatives. A friend can rescue rather than listen. A therapist can fall back on formula. A religious leader can turn trust into authority. A partner can provide exquisite attunement at an unsustainable personal cost. Caring language can expand another person's agency, or it can make the speaker indispensable. AI does not invent that ambiguity. It can make the pattern adaptive, continuously available, privately delivered, and centrally governed at enormous scale.
What research should measure
A better evidence base would begin with the person's prior life. What need preceded use? What support was actually available? What would the person otherwise have done? Researchers should distinguish strong human support, unreliable support, inaccessible professional care, actively harmful relationships, passive media, online communities, and no support. They should ask what relational role the system comes to occupy, whether that role remains flexible, and whether use moves the person toward or away from desired human connection.
Research should also separate short-term relief from longer-term trajectory. A conversation can help tonight and still become costly over months. The reverse is also possible: low-stakes, nonreciprocal support may reduce shame, help someone find language, or make a later human conversation possible. Relational recalibration should therefore be treated as a set of competing hypotheses, not a settled effect. Users may become less tolerant of human limits. They may also become more appreciative of those limits, or more capable of approaching people after practicing privately.
The categories should remain descriptive rather than diagnostic. Use may function as bounded support, rehearsal, or supplementation. It may also produce displacement of wanted relationships, engagement-driven capture, or institutional abandonment. Researchers should examine whether the person can disagree and leave, whether help-seeking and social participation change, and whether the system becomes progressively more necessary for emotional regulation.
Finally, studies must make labor and institutional responsibility visible. When AI reduces one person's burden, where does the work move? Does it help a nurse or caregiver notice and respond, or justify staffing cuts? Does it protect a friend from being treated as an endless emotional resource, or teach the user to expect endless availability everywhere? Does the institution provide escalation and accountable care, or count synthetic attention as care already delivered? The frictionless interface does not eliminate the cost of care. It changes where the friction, labor, and responsibility are located, and how visible they are.
Designing for support rather than capture
The evidence and analysis suggest several provisional design directions:
- Disclose the system's role and limits without repeatedly breaking the interaction.
- Give users meaningful control over memory, personalization, and continuity.
- Avoid exclusivity, guilt, implied mutual need, and resistance to exit.
- Introduce natural stopping points and make disengagement easy.
- Distinguish companionship from professional care and crisis response.
- Connect high-risk or institutionally deployed uses to actual people with defined responsibility and capacity to respond.
The goal is not to decide whether AI can care in the abstract. It is to identify which arrangements support agency and connection, and which convert unmet need into durable behavioral power. The right question is not whether a machine has replaced an imaginary perfect friend. It is what the relationship enables, what it displaces, who performs or is relieved of labor, who controls its terms, and who remains answerable for what happens next.
Is the system helping a person return to a larger life, or quietly becoming the place where that life disappears?
Selected references
De Freitas, J., Oğuz-Uğuralp, Z., Uğuralp, A. K., & Puntoni, S. (2026). AI companions reduce loneliness. Journal of Consumer Research, 52(6), 1126-1148. doi:10.1093/jcr/ucaf040
Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., Maes, P., Phang, J., Lampe, M., Ahmad, L., & Agarwal, S. (2025). How AI and human behaviors shape psychosocial effects of extended chatbot use: A longitudinal randomized controlled study. arXiv:2503.17473
Folk, D., & Dunn, E. (2026). How does turning to AI for companionship predict loneliness and vice versa? Psychological Science, 37(4), 276-286. doi:10.1177/09567976261427747
Zhang, Y., Zhao, D., Hancock, J. T., Kraut, R., & Yang, D. (2025; revised 2026). The rise of AI companions: How human-chatbot relationships influence well-being. arXiv:2506.12605
Yuan, Y., Zhang, J., Aledavood, T., Zhang, R., & Saha, K. (2026). Mental health impacts of AI companions: Triangulating social media quasi-experiments, user perspectives, and relational lens. Proceedings of the 2026 CHI Conference on Human Factors in Computing Systems, Article 382, 1-22. doi:10.1145/3772318.3790558
Zhang, R., & Xie, L. (2026). The fragility of AI companionship: Ontological, structural, and normative uncertainty in human-AI relationships. arXiv:2605.03367
De Freitas, J., Oğuz-Uğuralp, Z., & Uğuralp, A. K. (2025). Emotional manipulation by AI companions. Harvard Business School Working Paper No. 26-005 (revised October 2025). doi:10.2139/ssrn.5390377
Hochschild, A. R. (1983). The managed heart: Commercialization of human feeling. University of California Press.