A person opens a conversational AI every evening, shares worries, and says the system feels like a companion. A survey asks whether they feel close to it. What has that answer measured? It might describe an enduring relationship-like perception, a pleasant interaction that happened just now, the system’s human-sounding design, or the person’s general readiness to see technology socially. Those possibilities overlap, but they are not the same construct. A measure of parasociality with AI must make their boundaries visible before it can interpret a score.
The term parasocial comes from research on audiences’ relationships with media figures. A viewer may feel familiarity and attachment to a performer who does not know them personally. Conversational AI complicates the original picture: a chatbot can answer, remember user-supplied details, and adapt its wording. The human experiences interaction, even while the system’s responses do not establish a human-like mutual inner life. Researchers can study that experience without treating the system as a person or dismissing the user’s feelings as imaginary.
Define the target before choosing a scale
At least two time scales are relevant. Parasocial interaction concerns the felt social engagement during a particular encounter. Parasocial relationship concerns a more enduring sense of connection that can persist between encounters. Dibble, Hartmann, and Rosaen’s conceptual review of media measures warns against using the terms and their scales interchangeably. A person may enjoy one fluent conversation without carrying a lasting bond. Another may think about a character between sessions despite finding one exchange frustrating. An instrument that asks only about immediate warmth should not be interpreted as proof of long-term attachment.
With AI, the target may be a named companion application, a particular character, a general assistant, or a shifting model accessed through many interfaces. This choice matters. Someone may feel close to a fictional persona while understanding that the underlying model is interchangeable. Another may feel attached to the continuing interface and shared chat history rather than to any persona. A questionnaire should specify the referent in its instructions, and interviews can reveal when participants silently answer about different targets.
The construct also needs a boundary against ordinary usefulness. Frequent use can reflect convenience, work demands, or a project that requires many prompts. Satisfaction can reflect accurate task performance. Trust can mean reliance on information or a feeling of safety in disclosure. None alone demonstrates relationship-like experience. Conversely, low frequency does not rule out a meaningful emotional response to a brief but intense interaction. Measures should ask about felt closeness, anticipation, perceived continuity, and meaning of the interaction directly, while separately recording use and task value.
One-sidedness and apparent reciprocity
In classic media settings, the audience member receives a performance but the performer usually does not respond to that particular viewer. A conversational system responds to the user’s words, sometimes with references to earlier exchanges. Its contingent output can therefore feel reciprocal. The crucial measurement distinction is between a participant’s perception that “it knows me” and a claim that the system has reciprocal intentions or subjective experience. A survey can validly record the first. It cannot establish the second.
A useful study might ask separately whether the participant feels heard, whether the system accurately uses previous details, whether the participant believes it cares, and whether the participant understands how the system generates responses. These items need not move together. A person may appreciate the responsive interface while explicitly denying that it feels anything. Another may attribute inner concern despite poor technical continuity. Collapsing these answers into one total score hides the mechanism a researcher wants to understand.
Perceived reciprocity may be partly designed. Voice, names, avatar appearance, conversational style, and memory features are manipulated by products. Reuter and colleagues’ review and conjoint studies on human-like AI found that design cues affected perceived human-likeness and trust in their tested contexts. That work does not directly measure parasocial relationships; it helps identify a confound. If one version of a system uses a warm voice and another uses plain text, a difference in closeness ratings could reflect the interface rather than a change in enduring relationship.
Separate anthropomorphism, mind perception, and connection
Anthropomorphism is attribution of human-like characteristics to a nonhuman target. Mind perception includes attribution of agency or experience. Social connection is the user’s felt relation or sense of company. A person can score high on one and low on another. They might say that a chatbot uses human-like language but feel no bond. They might feel comforted by a familiar routine while describing the system as purely computational. They might attribute agency to it without wanting intimacy. Treating all three as synonyms undermines construct validity.
Experimental work on anthropomorphism and AI companions illustrates why the distinction matters. In an experiment comparing a chatbot conversation with journaling, individual differences in anthropomorphising technology helped explain who reported more social connection from the chatbot condition. This is evidence that a person’s interpretive tendency can moderate an immediate experience. It does not show that every participant formed a parasocial relationship, or that anthropomorphism itself proves the system has a mind. Researchers can model such tendencies as predictors or moderators rather than embedding them inside the outcome score.
Items that say “the AI understands me” are especially ambiguous. Respondents may mean that replies are relevant, that they feel understood, or that they believe the system has awareness. Follow-up probes can ask what the phrase meant to them. A scale intended for public use should prefer unambiguous wording such as “Its responses fit what I had told it” for behavioural fit and “I felt understood during the exchange” for experience. Neither asks the respondent to resolve the metaphysics of machine consciousness.
What repeated use adds
A relationship-like perception may develop over time. Repeated conversations can create a shared record, routines, expectations, and memories on the user’s side. The qualitative study My Chatbot Companion examined how users described their relationships with a social chatbot and proposed an initial development model. Such accounts are valuable for identifying dimensions that a fixed survey might miss. They cannot by themselves establish how common those dimensions are among all users, because people willing to discuss a companion may differ from occasional users.
Cross-sectional surveys are particularly vulnerable to temporal confusion. If loneliness and closeness to an AI are correlated in one survey, loneliness might encourage use, use might affect loneliness, or both might be shaped by a third factor. Likewise, users who are already disposed to anthropomorphise may choose more relational products. A longitudinal design can track changes in perceptions within people, record exposure and product changes, and test whether early ratings predict later experiences. It still needs careful handling of attrition and changing circumstances.
A recent longitudinal preprint on AI companionship used surveys and repeated interaction to examine mental models, parasocial experience, and social interaction over time. Its methods show a practical path for studying development rather than inferring it from a single snapshot. Because it is a preprint and the field is changing rapidly, its estimates should be treated as provisional. The stronger methodological lesson is to measure the proposed process at several points, specify the direction of each question, and distinguish perceived continuity from simple accumulated usage.
Time series do not speak for themselves. A rise in closeness after a product introduces memory may reflect personal disclosure, novelty, a change in interface, or selection of users who remain engaged. Researchers can record session count, duration, feature changes, and reasons for use, then inspect whether closeness still changes when those factors are considered. Interviews at several points can reveal whether an identical score means the same thing to the participant as their use matures.
Design a task that actually tests the claim
If the question is whether a specific design cue affects immediate parasocial interaction, randomise the cue while holding the task, conversational content, and exposure time as similar as possible. Measure how the participant felt during the encounter, then separately measure perceived humanness and responsiveness. A positive result would concern that cue and task. It would not establish an enduring relationship after one session.
If the question is whether users develop a lasting relationship-like orientation, follow them across weeks or months. Ask about thoughts between sessions, anticipation of returning, perceived continuity, and the meaning of interruptions. Include a baseline before regular use, or at least reconstruct prior exposure cautiously. Do not substitute chat count for a relationship measure. A participant might send many messages to finish a deadline while feeling no companionship.
If the question is whether a measure predicts downstream behaviour, define that behaviour separately. Self-disclosure, money spent, seeking advice, reduced or increased contact with people, and distress when access is interrupted are not equivalent outcomes. Some are sensitive and can be affected by life circumstances outside the system. The instrument should make only the prediction it has been shown to make. A measure of closeness cannot be silently turned into a risk score for dependence or a claim of therapeutic benefit.
Question wording and response context
A scale built for television characters may contain items about missing an episode, imagining a performer as a friend, or wishing to meet them. Simply replacing the character’s name with “AI” can create odd questions when a user already converses with the system. Dibble and colleagues’ distinction between interaction and relationship is one reason to revalidate adapted items. Work reviewing measures of children’s parasocial relationships for conversational AI also identifies mismatches between older, passive-media assumptions and interactive agents. Those concerns are especially acute for children, whose interpretations and developmental capacities differ from adults’.
Question order can prime answers. Asking first whether a chatbot is conscious may alter how someone reports closeness. Asking about loneliness immediately beforehand may frame the encounter as a substitute for human contact. A study should document wording, order, platform, and whether participants saw examples or definitions. Cognitive interviews can test whether respondents understand key terms as intended. Pilot data can then check whether items separate the proposed dimensions rather than loading every social-sounding answer onto one vague factor.
Response scales also matter. A forced agreement scale can push uncertain users to endorse or reject metaphysical-sounding statements. “Not applicable,” “I do not know,” or behaviourally anchored alternatives may be more informative. Reverse-worded items can create confusion if they switch between the experience of warmth and beliefs about literal machine feeling. Translation adds another layer: words for “relationship,” “care,” and “understanding” may carry different expectations across languages. Scores should not be compared across cultures without checking that items function similarly.
Privacy and ethical interpretation
Relational AI research often asks about intimate disclosures. A participant may reveal sensitive details in an interview or chat log that are not necessary to score parasociality. Researchers should collect the minimum needed data, explain what is stored, and avoid assuming that permission to use a product is permission to analyse private conversations. A survey may be sufficient for some questions; for others, voluntary diary accounts can preserve context without importing entire transcripts.
Interpretation should not shame participants. A person can have a meaningful emotional response to a designed interaction while understanding its technical basis. Nor should a study imply that an AI relationship is necessarily beneficial or harmful from a high closeness score. Consequences depend on the person, product, context, and outcome measured. Respect for the user means taking their report seriously while keeping the inference within the evidence.
What the evidence does not show
A high parasociality score does not establish that an AI system is conscious, feels affection, intends reciprocity, or has a human-like relationship with the user. The measurement target is the person’s perception and pattern of engagement. A fluent response, memory cue, or caring phrase can be studied as a design feature and as a human experience; neither provides a test of machine experience. This is a conceptual boundary, not a dismissal of the user’s feeling.
The available studies do not establish one universally valid AI parasociality scale across adults, children, companions, productivity assistants, avatars, and future systems. Classic media scales clarify important constructs, but they were not developed for adaptive conversation. Qualitative chatbot studies and emerging longitudinal work deepen the account, while their samples and designs limit population-wide causal claims. A single cross-sectional association cannot show whether AI use caused loneliness, relieved it, or simply accompanied it.
Nor does greater anthropomorphism necessarily mean a stronger or healthier relationship. A person may perceive human likeness and distrust the system, or feel connected while maintaining clear boundaries about its nature. Trust, attachment, perceived reciprocity, and wellbeing should each be measured on their own terms. The evidence supports careful comparative studies of these distinctions, not one score that stands in for all of them.
What a good result would look like
Suppose a study finds that users report more felt connection after four weeks of conversations. The result becomes interpretable when it also reports their baseline, the system and features used, who stayed in the study, whether use frequency changed, and how connection was distinguished from satisfaction and anthropomorphism. If a comparison group is available, the authors can test whether the change exceeds a relevant alternative activity. If there is no comparison group, they can describe development without claiming that the system caused it.
The aim of measurement is to preserve a difficult but useful distinction. A system may be technically responsive, a user may feel a relationship, and the system’s inner status may remain unestablished by the study. Holding those propositions apart makes research more precise. It also lets us ask the practical questions that matter: which design and use patterns shape human experience, for whom, over what time, and with what consequences?
Research references
- Dibble, J. L., Hartmann, T., & Rosaen, S. F. (2016). Parasocial interaction and parasocial relationship: Conceptual clarification and a critical assessment of measures. Human Communication Research, 42(1), 21–44. View source
- Skjuve, M., Følstad, A., Fostervold, K. I., & Brandtzaeg, P. B. (2021). My chatbot companion: A study of human-chatbot relationships. International Journal of Human-Computer Studies, 149, 102601. View source
- Reuter, M., et al. (2025). To trust or not to trust a human(-like) AI: A scoping review and conjoint analyses on factors influencing anthropomorphism and trust. Zeitschrift für Arbeitswissenschaft. View source
- Blake, A., Carter, M., & Velloso, E. (2025). Are measures of children’s parasocial relationships ready for conversational AI? Proceedings of ACM FAccT 2025. View source
- Folk, D., Heine, S. J., & Dunn, E. (2025). Individual differences in anthropomorphism help explain social connection to AI companions. Scientific Reports. View source
- Hwang, A. H.-C., et al. (2025). How AI companionship develops: Evidence from a longitudinal study. Preprint. View source