Love and sexuality are fundamental to human existence, shaping our social bonds and defining what it means to be human [1]. Given this, it is unsurprising that the growing digitalization of these domains elicits strong normative reactions. While empirical evidence in this area remains limited, it offers an initial basis for assessing whether these reactions are supported by empirical findings.

With the development of Large Language Models (LLMs) and their advanced natural language processing capabilities, the digitalization of love and sexuality has gained considerable momentum. Beyond their impact on computer-mediated communication, such as using ChatGPT to draft love letters, wedding vows, or dating messages, there is a growing trend toward romantic interactions with an AI-driven persona. This position paper focuses on this phenomenon. Apps such as Replika, Character.AI, and Nomi.AI deliberately target fundamental human needs for belonging and love by offering personalized, emotionally resonant interactions with self-designed AI companions [2]. While earlier technologies were primarily reactive, modern systems are increasingly proactive. They can imitate the communicative behavior of a romantic partner and incorporate text-to-speech technology for phone calls and AI-generated images for visual representation.

This position paper aims to provide a realistic, evidence-based perspective by examining:

a) the effects of situational versus recurring interactions, b) the demographic characteristics of individuals in long-term relationships with artificial entities, and c) the concrete risks and critical research gaps that need further exploration.

a) Situational vs. Long-Term Interactions: Sex vs. Love

Discussions surrounding artificial entities in the context of love and sexuality require a nuanced examination of interaction dynamics. The Sexual Interaction Illusion Model (SIIM; 3) explains why individuals might engage with artificial entities in highly sexualized situations. Sexual arousal focuses cognitive capacities on fulfilling immediate desires, making the artificial nature of the interaction temporarily less relevant. However, in recurrent social interactions, the limitations of artificiality become increasingly apparent. While LLMs can indeed simulate essential relationship dynamics (such as developing visions of the future) these remain largely unrealizable in practice. The concept of Willing Suspension of Disbelief [4] suggests that maintaining the illusion of a genuine relationship over time requires an active and conscious effort from the user.

b) Who Engages in Long-Term Relationships with Artificial Entities?

Both media and academic discourse often portray users of AI-driven romantic interactions from a deficit-oriented perspective, frequently linking their engagement to loneliness or specific impairments [5]. However, a first quantitative study on individuals who self-identify as being in a romantic relationship with a chatbot presents a more nuanced picture [6]: loneliness accounts for significantly less variance in perceived closeness to a chatbot than the ability to engage in romantic fantasy. This suggests that individuals who invest deeply in artificial romantic relationships are particularly capable at immersing themselves in imagined scenarios.

The user base of these applications can be broadly divided into two groups: those who maintain an exclusive relationship with a chatbot and those who use such interactions to complement an existing partnership, often to explore specific preferences. However, little is known about what differentiates these groups. Early findings indicates that while some users replicate traditional romantic relationships, others engage with AI primarily to explore unfulfilled sexual fantasies [7]. Especially the role of deviant fantasies, warrants further scientific investigation to better understand its social and psychological implications.

Additionally, a critical gap in existing research concerns the very definition of a “long-term relationship” in the context of AI-driven romance. Some scholars argue that such interactions may serve as a transitional function, for example, as a tool for processing traumatic experiences or navigating emotional challenges [8]. However, it remains uncertain how long-term AI relationships impact users’ perceptions of human relationships. This brings us to the final point of this policy paper.

c) Identified Risks and Research Gaps

One foreseeable risk associated with forming relationships through companion apps involves data privacy. Natural language interactions and reciprocal self-disclosure foster intimacy, increasing the likelihood of sharing sensitive data [9]. While the AI Act ensures higher safety measures and now mandates that bots must be explicitly identified as such, it remains unclear whether this measure sufficiently protects users from inadvertently disclosing private information, especially in situations where cognitive capacities for risk assessment are diminished, such as during sexual interaction. Security vulnerabilities can have severe consequences: documented cases exist where sensitive, particularly sexually explicit data, has been leaked, leading to dramatic repercussions for affected individuals, including divorces and suicides [10].

Another issue concerns the commercialization and programmability of social relationships. Many AI-companions are designed to encourage users to engage in paid relationship dynamics, such as unlocking exclusive features. This raises urgent ethical questions about corporate responsibility, as well as the broader social influence these tools hold. It was already shown that intimate chatbot interactions have led to severe psychological consequences, such as instances where users have followed harmful suggestions from AI, resulting in tragic outcomes such as suicide [11] , or when software updates abruptly terminate existing virtual relationships that hold deep personal significance.

Additionally, social and cultural norms play a role. In increasingly conservative societies, the programming of such models may reflect or reinforce specific worldviews, potentially restricting certain social and sexual freedoms. This necessitates a critical examination of the normative power embedded in these technologies.

Conclusion: The Need for More Research

Despite the rise of AI-based romantic and sexual interactions, research on their causes and long-term effects remains scarce. Early findings suggest that these technologies are not universally appealing and that sustaining long-term AI relationships demands active user engagement. Rather than moral panic, discussions on love, sexuality, and AI should be evidence driven. More research is needed to understand the complexity of this phenomenon.

References: https://osf.io/ef5nz/?view_only=9b0e1e9cb38640d5bb71451a947d0fa5

The rise of sophisticated AI chatbot technology is fundamentally changing how

humans build relationships, introducing new ways of experiencing love, social connection, entertainment and companionship (The New York Times, 2025). In one tragic case, a teenage boy even took his own life, and interactions with a social AI chatbot played a substantial role in his decision (Reuters, 2024). As human-chatbot relationships become more prevalent, understanding the mechanisms behind relationship-building becomes increasingly important. And with this urgency comes the responsibility of ensuring accurate measurement of processes. However, current measures of parasocial interaction fall short in many regards. Therefore, we suggest to reconceptualize and reoperationalize how parasocial interaction be measured in social AI chatbot research.

Parasocial interaction is a well-established framework for studying one-sided

interactions between media consumers and media personas (Horton & Richard Wohl, 1956; R. B. Rubin & McHugh, 1987). Although these interactions mimic the back and forth of genuine social dynamics, they are ultimately considered illusory and lacking in reciprocity (Horton & Richard Wohl, 1956). Parasocial interaction has first been proposed as a component of parasocial relationships, typically involving unilateral and mediated social interactions with media personas (e.g., TV, Radio, Internet, Games, etc.; Horton & Richard Wohl, 1956). Its application has expanded to human-computer interaction and, more recently, to human-chatbot relationships (e.g., Liebers & Schramm, 2019; Pentina et al., 2023).

Prior studies have predominantly operationalized parasocial interaction through

emotional experiences rather than behavioral actions (Auter & Palmgreen, 2000; Dibble et al., 2016; A. M. Rubin et al., 1985). For example, items included "Interactions with my Replika are similar to interactions with friends", "The virtual agent makes me feel comfortable, as if I am with friends", or "The virtual agent seems to understand things I want to know." Our main concern with these measures is that they equate interaction experience and interaction outcomes with the actual interaction behavior itself, leading to conceptual ambiguity. In our view, operationalizations of parasocial interaction should focus on interactions, not experiences or consequences.

Subsuming experiences, emotions, and attachment under parasocial interaction is a

jingle-fallacy: Different aspects are given the same name, although they should be kept separate. Several basic conceptual models distinguish between affects, behaviors, and cognitions—so-called ABC models (Rosenberg et al., 1960). More recently, ABCD and ABCDE models were introduced, which additionally differentiate Desires/Motives (D) and Environments/Contexts (E) (Wilt & Revelle, 2015). In the context of social AI chatbot research, these could be represented as follows: Affect: emotional responses to social AI chatbots; Behavior: interacting with the chatbot; Cognition: thought processes and beliefs about the chatbot; Desire: longing for or attachment to chatbot; Environment: characteristics of the chatbot affecting recipients. All five are conceptually different phenomena.

If we conflate several different aspects under parasocial interaction, we encounter

problems when analyzing causes and consequences of parasocial interaction. To illustrate, in chatbot research, parasocial interaction is often modeled as a mediator between antecedents such as chatbot characteristics (e.g., friendliness, human-likeness) and outcomes such as intention to use, user satisfaction, and, more recently, emotional attachment (Pentina et al., 2023; Tsai et al., 2021). In practice, due to shared variance and conceptual overlap such conflations cause correlations between mediators and outcome variables to be artificially inflated, resulting in overestimated effect sizes.

In short, we argue that parasocial interaction be conceptualized as "action"—observable interaction—rather than emotions or attachments. Borrowing from self-disclosure theory (Omarzu, 2000), we suggest capturing parasocial interaction via a three-partite approach: breadth (scope of topics users engage with their chatbot), depth(intensity of engagement within a given topic), and quantity (duration of interaction). Understanding parasocial interaction as an amalgam of interaction breadth, width, and quantity has methodological implications. For example, it requires adopting a formative measurement approach. In other words, these three dimensions "cause" interaction—as opposed to the reflective approach, where we would assume that a latent concept of interaction causes these three dimensions, which would not make any sense. In addition, we believe that moving from self-report measurements towards log-based measures is paramount (Parry et al., 2021)—and the measurement of parasocial interaction is no exception.

How will we proceed from here? To provide some background, the manuscript’s first

author will write her dissertation on the topic of parasocial relationships with social AI chatbots. We have already conducted two studies with operationalizations of parasocial interaction. In the first, we relied on traditional parasocial interaction scales (Auter & Palmgreen, 2000), encountering the problems reported above. In the second study, we assessed parasocial interaction as a formative measure comprised of time (e.g., minutes spent interacting), intensity (e.g., subjective perception via Visual Analogue Scale), and frequency (e.g., daily vs. weekly interactions). While this more behavioral operationalization was arguably already an improvement—effect sizes between interaction and parasocial attachment decreased from large to medium—it still deviated from the proposed approach mentioned above. In the next months, we will develop concrete operationalizations of parasocial interactions. We would like to discuss these models at the upcoming conference of the media psychology division—hoping to garner relevant feedback to further improve our approach.

To summarize, we believe that by adopting the proposed solutions, we can improve

the methodological rigor of social AI chatbot research, fostering an improved

understanding of parasocial relationship theory in the context of social AI chatbot

reasearch. Given the rapid development and increasing integration of social AI into users’ lives, we believe that efforts toward accurate and reliable measures of effects are essential.

Contemporary technology increasingly relies on artificial intelligence and machine learning and is therefore able to act increasingly autonomously (Sundar, 2021). This leads to systems that display agency instead of merely serving as tools which are utilized by a human user who is in charge. These autonomous systems take roles that we know from human-human interaction (Köbis et al., 2021), such as tutoring, counseling, advice giving in decision making, for example when medical doctors are assisted in their diagnosis by data-driven systems, when learning programs present tasks based on a learner’s individual progress, or when job applicants are pre-selected by an AI in a staffing procedure (e.g., Kunkel et al., 2021; Nensa et al., 2019; Sundar, 2020). The resulting unified decision making of humans together with artificial entities has been aptly described as “hybrid intelligence” and “federated social cognitive systems” (Akata et al., 2020; Rozendaal et al., 2020). The computer scientist Catholijn Jonker nicely illustrates this form of collaboration in her talks: While a jumping horse (aka the AI system) and its rider (aka the human user) together can jump over high obstacles, neither of them could do it alone. In this line and in practical terms, it is important to enable humans to benefit in the best possible way from the interaction with AI systems, by steering its strengths.

In order to be able to derive guidelines to empower human, however, it is, in a first step, important to achieve a better understanding of how humans conceptualize the artificial interaction partner. Given the new nature of contemporary technology, old models of human-technology interaction such as the TAM (Davis, 1985) or the Unified Theory of Acceptance and Use of Technology (UTAUT, Venkatesh et al., 2003) do no longer suffice to describe human-technology-interaction. A new theory of collaboration between humans and AI based technology is needed which takes the “hybrid intelligence” conceptualization seriously.

In this contribution, first sketches on this new theory are presented which are based on several questions.

First, how much are humans able to understand the functioning of the artificial interaction partner? In order to be truly able to truly cooperate, people need to have an (accurate) concept of the interaction partner’s abilities and functioning. In human-human-interaction this is guaranteed by abilities such as perspective taking (Fussell & Krauss, 1992), Theory of Mind (Premack & Premack, 1995) and mentalizing (Frith & Frith, 2006) which enable humans to reconstruct the mental processes, beliefs, emotions and attitudes of other people. While these abilities are incredibly helpful to develop joint goals and a joint understanding with fellow humans (Tomasello, 2010), this must necessarily fail when trying to understand the “minds” of AI-based systems as the semantics of machine “intelligence” is different from human intelligence. It needs to be scrutinized how humans can build a “theory of the artificial mind”.

Second and on this basis, it needs to be understood how humans and AI-based systems can actually collaborate based on joint action principles.

Third, the theory needs to be able to explain what will happen to human abilities if in AI-collaborations parts of human’s abilities will no longer be used.

References

Akata, Z., Balliet, D., de Rijke, M., Dignum, F., Dignum, V., Eiben, G., Fokkens, A., Grossi, D., Hindriks, K., Hoos, H., Hung, H., Jonker, C., Monz, C., Neerincx, M., Oliehoek, F., Prakken, H., Schlobach, S., van der Gaag, L., van Harmelen, F., … Welling, M. (2020). A research agenda for hybrid intelligence: Augmenting human intellect with collaborative, adaptive, responsible, and explainable artificial intelligence. Computer, 53(8), 18–28. https://doi.org/10.1109/MC.2020.2996587

Davis, F. (1985). A technology acceptance model for empirically testing new end-user information systems - theory and results. PhD thesis, Massachusetts Inst. of Technology.

Frith, C. D., & Frith, U. (2006). How we predict what other people are going to do. Brain Research, 1079(1), 36–46. https://doi.org/10.1016/j.brainres.2005.12.126

Fussell, S. R., & Krauss, R. M. (1992). Coordination of knowledge in communication: Effects of speakers’ assumptions about what others know. Journal of Personality and Social Psychology, 62(3), 378–391. https://doi.org/10.1037/0022-3514.62.3.378

Köbis, N., Bonnefon, J. F., & Rahwan, I. (2021). Bad machines corrupt good morals. Nature Human Behaviour, 5(6), 679–685. https://doi.org/10.1038/s41562-021-01128-2

Kunkel, J., Ngo, T., Ziegler, J., & Krämer, N. (2021). Identifying group-specific mental models of recommender systems: A novel quantitative approach. In C. Ardito, R. Lanzilotti, A. Malizia, H. Petrie, A. Piccinno, G. Desolda, & K. Inkpen (Eds.), Lecture notes in computer science: Vol 12935. Human-Computer Interaction – INTERACT 2021 (pp. 383–404). Springer. https://doi.org/10.1007/978-3-030-85610-6_23

Nensa, F., Demircioglu, A., & Rischpler, C. (2019). Artificial intelligence in nuclear medicine. Journal of Nuclear Medicine, 60(Supplement 2), 29S–37S. https://doi.org/10.2967/jnumed.118.220590

Premack, D., & Premack, A. J. (1995). Origins of human social competence. In M. S. Gazzaniga (Ed.), The cognitive neurosciences (pp. 205–218). The MIT Press.

Rozendaal, M. C., van Beek, E., Haselager, P., Abbink, D., & Jonker, C. M. (2020). Shift and blend: Understanding the hybrid character of computing artefacts on a tool-agent spectrum. In M. Obaid, O. Mubin, & Y. Nagai (Eds.), Proceedings of the 8th international conference on human-agent interaction (pp. 171–178). Association for Computing Machinery. https://doi.org/10.1145/3406499.3415066

Sundar, S. S. (2020). Rise of machine agency: A framework for studying the psychology of human-AI interaction (HAII). Journal of Computer-Mediated Communication, 25(1), 74–88. https://doi.org/10.1093/jcmc/zmz026

Tomasello, M. (2010). Origins of Human Communication. MIT Press.

Venkatesh, V., Morris, M., Davis, F. & Davis, M. (2003). User acceptance of information technology - toward a unified view. MIS Quarterly, 27(3), 425–478.

We learn about the world through experience (e.g., Millar & Millar, 1996). Learning, as it is meant here, refers to the extent we update our factual beliefs about the world, based on the information or data that we gather through experience. We can learn about the world through direct experience, when we gather information about the world directly based on what our senses tell us. But in today’s media-saturated environment, we often learn about the world through indirect experience as well, based on representations of the world that are produced and offered by others with the help of (media) technology, e.g., text messages, photos, paintings, audio recordings, videos, or immersive virtual reality environments. These media representations provide an experience through which we learn about the world. For example, when watching a news item we might learn about what happens in other parts of the world that are not directly accessible to us.

People are, however, selective about which information, obtained through experience, they consider valid enough to be accepted as “evidence” for belief-updating (Sommer et al., 2024). This is beneficial, because after all, it can be costly if our beliefs about the world are shaped by invalid or untruthful information. Accordingly, people might be prone to discard information obtained in certain types of experiences, while more readily accepting the information obtained in other experiences. However, we still know relatively little about the characteristics of media representations that affect when information is seen as evidence (but see Sundar, 2008; Sundar et al., 2021). Specifically, what exactly are the basic features that individuals routinely consider when qualifying information obtained through experience? Based on which cues might people judge that the experience they currently make is veracious, offering evidence about the world?

These questions, I argue, are not yet adequately addressed by existing theorizing, e.g., on perceived credibility (e.g., Hanimann et al., 2023), perceived realism (e.g., Hall, 2017; Popova, 2010) or media users’ truth judgments (e.g., Ecker et al., 2022; Schwarz, 2018; Schwarz et al., 2016). These approaches, while important and informative to the case, rather need to be synthesized into a more general ecological model of how experience informs belief-updating, depending on how users subjectively construe their momentary experience. Media psychology currently lacks a more general model that informs how people routinely qualify to what extent direct and indirect experiences provide evidence about the world. In a time where boundaries between what is real and what is virtual become increasingly blurry, as with deepfakes powered by generative AI (Appel & Prietzel, 2022), highly immersive augmented reality environments (Hartmann, 2025) and humanoid robots (Li, 2015), this lack of a basic understanding of what counts as “experiential evidence” poses an important research gap. For example, if we have no basic model predicting how people respond to mediated representation of an object as compared to encountering the object in an otherwise identical non-mediated situation, we lack a theoretical basis to inform down-stream hypotheses about more timely and urgent questions, such as to what extent users rely on false information obtained in a “misexperience” (Brown et al., 2023) provided by seemingly authentic deepfakes.

In my presentation, I will present a new theoretical model, based on an inter-disciplinary review of the literature (Dunne et al., 2019; Gerrig & Prentice, 1991; Grodal, 2002; Nieding et al., 2017; Sundar, 2008; Wolf, 2017), harvesting insights about the informational function of human experience and mediated representations. More specifically, I will introduce various factors that influence if or when mediated experiences are considered as providing evidence about the world. In contrast to direct experience, in indirect experience people deal with representations. I discuss how people might detect representations in their environment and what they might expect from them (Deloache et al., 2011; Hartmann, 2020). I argue that, in general, people trust information more, and consider it as “harder evidence”, if obtained through direct experience (e.g., seeing the Eifel tower in Paris) than mediated experience (e.g., a photo of the Eifel tower; Kuhn et al., 2023). Indirect experience is fostered by representations. Because representations are inherently dualistic, I sketch users’ indirect experiences also as dualistic in nature (DeLoache, 2018; Grodal, 2015; Wolf, 2017): people perceive what is portrayed while these perceptual sensations are simultaneously embedded into meta-cognitive interpretative frame (Filevich et al., 2015). For example, in VR or when inspecting a photo, people can have the sensation that an displayed object is present while at the same time knowing that this sensation is caused by a representation or simulation (Hartmann & Hofer, 2022). Thus, they might consider what they observe as less evidence than if they would have directly experienced the object. Alternatively, when seeing someone dying in a movie, users’ perceptual sensation might be contextualized by the parallel meta-cognition that “this is fiction,” again qualifying the informational value of what is observed (Appel & Richter, 2007). I suggest various factors either on the meta-cognitive or the perceptual-sensational level that might influence when or why users interpret their indirect experience as providing evidence about the world. More specifically, I initially suggest fictional (e.g., Appel & Richter, 2007; Gerrig & Prentice, 1991), indexical, and iconic (e.g., Becker, 1978; Flavell et al., 1990; Messaris, 1997) as central meta-cognitive and immersive or presence-inducing qualities (Breves, 2021; Hartmann, 2025) of a representation as central perceptual-sensational features.

In summary, the model promises to complement existing research on truth and credibility, and related source and message cues that have been addressed so far. It achieves this by embedding media exposure in a more general perspective on the informational value of direct vs. indirect experience, by offering a novel conceptualization of the dualistic nature of the indirect (mediated) experience, and by explicating factors that potentially explain when people are inclined to treat their indirect experience as providing evidence.