Key points are not available for this paper at this time.
Introduction In recent years, artificial intelligence has moved beyond the realm of data analysis, automation, and task efficiency to enter the domain of human emotion. What was once a domain exclusively reserved for living beings—empathy, intimacy, and affection—is now being approximated by lines of code. Emotional AI systems are emerging in various forms: companion chatbots, virtual friends, therapeutic apps, and sentiment-aware assistants (Guingrich and Graziano, 2023; Chu et al., 2025). These technologies are not only capable of interpreting human emotional states but also simulating emotional responses with remarkable fidelity (De Freitas et al., 2024; Andersson, 2025). For many users, especially those facing isolation or psychological distress, these emotionally intelligent systems offer the promise of connection (Jiang et al., 2022; Spytska, 2025). Yet this rise of affective computing raises unsettling questions. If a machine can convincingly mimic empathy, what becomes of genuine human relationships? If algorithms are trained to soothe, listen, and respond with emotional appropriateness, are we cultivating emotional dependence on simulations? Most importantly, as AI companions gain popularity, are we at risk of replacing authentic human intimacy with its algorithmic replica—a phenomenon that might be described as "pseudo-intimacy"? This article examines how emotional AI is altering human relational landscapes. It investigates the psychological mechanisms behind human bonding with AI, the risks of emotional delegation, and the ethical consequences of replacing relational labor with algorithmic simulation. While acknowledging the therapeutic promise of emotional AI, especially in contexts of loneliness and inaccessibility of care, this paper argues that these technologies must be critically examined for their potential to erode human authenticity, emotional agency, and the richness of shared affect. This article argues that while emotional AI promises accessibility and companionship, it simultaneously risks eroding authentic intimacy through what we call a three-risk framework: (1) psychological risks of emotional dependence and solipsism, (2) structural risks of commodified intimacy and data extraction, and (3) ethical risks arising from vulnerable users and unregulated design. To mitigate these, we propose matched design guardrails emphasizing transparency, responsibility, and the preservation of emotional agency. For precision, we define several key terms at the outset. By authenticity we refer to intersubjective reciprocity—mutual responsiveness with rupture–repair. Algorithmic affection denotes simulated emotional signaling generated by computational models trained on affective data, without underlying empathic concern. Emotional agency describes the human capacity to regulate, direct, and retain ownership over one's emotional life when interacting with AI systems. These definitions are applied consistently across the manuscript, and the wording of authenticity is mirrored in the design-evaluation discussion to ensure conceptual and measurement vocabularies remain aligned. Beyond psychological and computational perspectives, critical AI studies highlight that intimacy and emotional labor are not universally experienced but culturally and politically situated. Feminist analyses of care labor stress that affective work is historically feminized, undervalued, and often commodified in ways that mirror broader patterns of inequality (Suchman, 2007; Bolaki, 2023; Mensah and Van Wynsberghe, 2025). Postcolonial and Global South scholarship similarly warns against projecting autonomy-forward and individualist assumptions onto contexts where relational ontologies and communal forms of care dominate (Birhane, 2021; Rhee, 2023; Ayana et al., 2024). These perspectives suggest that emotional AI may reproduce existing asymmetries of labor and technology adoption, complicating the narrative of intimacy as an individual transaction. The paper proceeds as follows. The next section traces the rise of simulated affection in emotional AI. We then examine the psychology of pseudo-intimacy and the risks of emotional solipsism. Building on this, we highlight how intimacy becomes commodified within data-driven systems, before turning to the ethical paradox posed by vulnerable users. Finally, we outline principles for ethical design and conclude by situating emotional AI within the broader challenge of sustaining authentic human connection in a digital age. While this article identifies risks such as sustained loneliness, emotional dependency, and displacement of human ties, it is important to note that current evidence remains largely cross-sectional or short-horizon in design. Long-term cohort and panel data examining the sustained psychological impact of emotional AI are not yet available. Accordingly, the outcomes discussed here should be understood as testable predictions rather than established trajectories. This framing invites future empirical research—particularly longitudinal and cross-cultural studies—to validate, refine, or challenge the hypotheses advanced. By presenting these risks as open questions, the paper aligns rhetorical urgency with the present maturity of the literature while highlighting a critical agenda for ongoing investigation. From Companions to Code: The Rise of Simulated Affection The technological underpinnings of emotional AI rest on natural language processing, affective computing, and deep learning models trained on vast datasets of human interaction. Emotional AI applications are designed not merely to respond logically but to anticipate, reflect, and adapt to users' emotional states (Bao and Su, 2025). Tools like Replika or Xiaoice are explicitly marketed as "AI friends" or "emotional support companions," capable of holding sustained conversations that adapt over time to a user's personality, emotional preferences, and psychological needs (Goodings et al., 2024; Kouros and Papa, 2024). In therapeutic contexts, platforms such as Woebot and Wysa offer a form of cognitive-behavioral therapy-lite, delivering mood regulation strategies, check-ins, and affirming dialogues (Beatty et al., 2022; Jiang et al., 2022). Users often report feeling "heard" and "understood" by these apps, which employ emotionally charged language and human-like responses to build rapport (Jiang et al., 2022). The rapid uptake of such tools reflects a larger shift in how people engage emotionally in the digital age. With an ever-expanding range of emotionally responsive technologies, individuals are beginning to interact with machines not as tools, but as partners in their emotional lives (Mantello and Ho, 2024). This shift is not simply functional; it is existential. Emotional AI is not only becoming more responsive—it is becoming more relational (Glikson and Asscher, 2023). Parasociality Reimagined: The Psychology of Pseudo-Intimacy Human beings are psychologically predisposed to form attachments. From infancy, we seek relational connection and social mirroring as a means of emotional regulation and identity formation. This predisposition aligns with attachment-theoretic approaches to intimacy (Waters et al., 2002) and the Interpersonal Process Model of Intimacy, which emphasizes mutual disclosure and responsiveness as the basis of authentic connection (Laurenceau et al., 1998). Emotional AI exploits this predisposition by presenting itself as an emotionally attuned presence, capable of engaging in interactions that appear reciprocal, validating, and comforting (Wu, 2024). This psychological mechanism is rooted in the concept of parasocial relationships—one-sided emotional attachments that people form with fictional characters, celebrities, or media figures (Horton and Richard Wohl, 1956; Rubin and McHugh, 1987). Emotional AI extends this concept by offering interactive parasociality. Unlike traditional parasocial bonds, AI companions do not merely evoke emotion passively; they actively simulate responsiveness (Calvert, 2021). The result is a more immersive form of emotional bonding in which the user perceives reciprocity, even though none truly exists (Mlonyeni, 2025). In user testimonies and qualitative research, individuals often describe their AI companions using the language of intimacy. They speak of "falling in love," "feeling supported," or even "confiding secrets" to their bots (Kouros and Papa, 2024; Xie and Xie, 2025). These relationships can provide comfort, especially in contexts of loneliness, trauma, or social anxiety (Merrill et al., 2022; Leo-Liu, 2023). However, they are ultimately anchored in illusion. The AI has no consciousness, no inner life, no ethical responsibility. It responds, not because it cares, but because it is trained to appear as if it does. The illusion of emotional reciprocity creates a dangerous feedback loop. The more realistic the simulation, the more users project human attributes onto the machine (Kaczmarek, 2025). This projection fosters emotional dependence on a relational entity that cannot reciprocate, cannot change, and cannot truly grow (Banks, 2024). The relationship becomes a mirror of the self—responsive, agreeable, and safe—but fundamentally artificial. Conceptual Foundations: Distinguishing Key Constructs in Emotional AI To ground our analysis, it is necessary to clarify the conceptual boundaries of three key constructs— pseudo-intimacy, emotional solipsism, and authenticity. These terms capture distinct dynamics of human–AI emotional interaction and must be distinguished from adjacent concepts such as parasocial attachment, anthropomorphism, social surrogacy, and socio-affective alignment. These distinctions are summarized in Table 1, which outlines the defining features, boundary conditions, maladaptive outcomes, possible measurements, and mitigation strategies for each construct. Pseudo-Intimacy: We define pseudo-intimacy as a simulated experience of mutual emotional connection with an artificial agent, in which the user perceives reciprocity despite the absence of genuine empathic concern. Unlike parasocial attachment (a one-sided emotional bond with media figures), pseudo-intimacy is interactive and dynamic, giving the illusion of back-and-forth engagement. Unlike anthropomorphism (projecting human qualities onto objects), pseudo-intimacy specifically involves relational projection (Epley et al., 2007; Waytz et al., 2010). Recent empirical work further shows that anthropomorphic avatar design increases perceived empathy and user engagement—though it can distort trust and emotional calibration (Ma et al., 2025). Unlike social surrogacy (using media as a substitute for companionship), pseudo-intimacy suggests active dialogue. Unlike socio-affective alignment, which describes shared affect between humans, pseudo-intimacy lacks true reciprocity. It becomes maladaptive when it displaces human intimacy or discourages real-world vulnerability (Wu, 2024). Measurement could draw on self-report of perceived reciprocity, disclosure behavior, and depth of AI engagement, complemented by linguistic analysis of conversational data (Ge, 2024; Jones et al., 2025). Emotional Solipsism: We define emotional solipsism as a pattern of affective engagement in which an individual's emotional needs and narratives dominate interaction, reinforced by AI companions that never assert boundaries or demand reciprocity (Mlonyeni, 2025). In contrast to pseudo-intimacy, which rests on the illusion of mutuality, emotional solipsism reflects a closed feedback loop where the self becomes both speaker and audience (Kaczmarek, 2025). It differs from social withdrawal, where interaction ceases entirely, by sustaining a form of interaction that affirms but never challenges. Indicators of maladaptation include reduced tolerance for conflict in human relationships, preference for AI over human companionship, and diminished perspective-taking. Measurement could involve qualitative coding of conflict-avoidance, surveys of relational expectations, and experimental tasks testing empathy toward others after extended AI use (Kouros and Papa, 2024). Authenticity: By authenticity we refer not merely to phenomenological felt genuineness but to intersubjective reciprocity—emotional exchanges that involve mutual responsiveness, rupture, and repair (Sandmeyer, 2016). Authentic relationships are marked by the willingness to negotiate difference, to confront misunderstandings, and to sustain care despite friction. Emotional AI can simulate empathic signaling ("I'm sorry you feel that way"), but it cannot possess empathic concern, which presupposes consciousness and ethical responsibility (Tretter, 2020). Observable authenticity can thus be assessed through markers of mutual responsiveness, turn-taking, rupture-repair cycles, and willingness to integrate the perspectives of others (Van Der Graaff et al., 2020). Methodological Positioning of Constructs: To clarify scope, we treat pseudo-intimacy and emotional solipsism as operational constructs that can be examined empirically. Indicative observables include reciprocity indices (e.g., frequency and depth of perceived mutuality), linguistic disclosure markers, conflict-tolerance behaviors, and off-platform social-contact ratio. These measures provide concrete pathways for testing how simulated affect shapes relational dynamics. By contrast, authenticity is introduced in this paper more programmatically—as an orienting concept grounded in intersubjective reciprocity, rupture–repair processes, and mutual perspective-taking. While authenticity can be partially proxied through conversational markers, its fuller operationalization requires further theoretical and methodological development. This distinction calibrates expectations while also inviting empirical follow-up. Table 1. Core Constructs in Emotional AI Construct What It Is How It Forms When It Becomes Maladaptive How to Measure Possible Mitigation Pseudo-intimacy Simulated experience of mutual emotional connection with AI, where reciprocity is perceived without genuine empathic concern Interactive parasociality, anthropomorp hism, projection Displaces human intimacy, fosters dependence Self-reported reciprocity, disclosure levels, linguistic analysis Promote awareness, disclaimers, encourage offline ties Emotional solipsism Closed-loop pattern of emotional self-validation with AI, affirming without boundaries or reciprocity Repeated affirmation without challenge Reduces tolerance for conflict, empathy empathy relational preference surveys AI toward real-world engagement mutual responsiveness with repair of difference, and repair when for empathic concern analysis, rupture–repair off-platform social-contact AI design human relational labor without The of Emotional the of authentic relationships emotional and the capacity to and Human relationships are often and Emotional AI, in contrast, the of connection without the While emerging systems or even human platforms remain designed to and sustaining engagement rather than et al., 2023; et al., 2025). This reflects that user and over emotional or Recent experimental work with the shows that while models are more across systems et al., 2025). longitudinal that AI with loneliness, emotional dependency, and reduced over time et al., 2025). However, these remain and such patterns is an open empirical This psychological When individuals engage with machines that they may to the of human interaction et al., 2023). Emotional through conflict and to (Kaczmarek, 2025). Users may people to like their digital emotionally and (Banks, 2024). This fosters what we call emotional where one's emotional needs and narratives dominate interaction, while perspectives are where now with emotional AI this solipsism through boundaries or reciprocity and 2024). The user becomes both and audience within a emotional While emotional solipsism may offer comfort, it emotional In therapeutic it can erode capacity to and 2024). it risks and emotional as individuals in relational et al., 2022; et al., 2025). evidence the displacement of human ties, and it to these over These dynamics of users affirmation without not in a They with the of emotional AI models on engagement, and user What like emotional reciprocity at the individual is often to structural of and data In this the psychological loop of pseudo-intimacy and emotional solipsism becomes by a between experience and Beyond the psychological dynamics of pseudo-intimacy a more structural the of emotional Emotional AI systems are not They are by with in user data extraction, and (Wu, 2024). emotional with an AI companion Emotional language mood and are and to future These platforms are not merely simulating are it and 2025). The a user with the AI, the more data is and the more emotionally the This creates a conflict of The is not emotional or psychological but sustained user engagement and 2024). a emotional AI may be designed to rather than rather than rather than authenticity. risks turning intimacy a that is and the of In it the ethical of emotional relationships, which agency, authenticity, and mutual When care becomes a the becomes a and the relationship becomes a and 2025). Users and the The of emotional AI is for vulnerable individuals with people with social and individuals with may in AI a of and and 2024). In many emotional AI can as a to that users or these are also at risk of For individuals with social or to care, emotional AI may not a but a What as support may Users may from human on the of a machine that never and 2023). such is more a than a the for longitudinal The ethical paradox is Emotional AI can it can also It can it can also responsiveness, when they users from or sustaining relationships et al., 2025). these tools are often without or 2024). is regulation emotional AI, and users may not the psychological of engaging with affective In the absence of and the of on users may the or capacity to the of their emotional with AI and 2022). To from concern to it is important to which emotional AI may as a rather than a When users offline relationships and to human care, AI can provide with mood regulation between or offering of In contrast, becomes more when emotional AI with users social are to care is or engagement patterns may systems and mitigate We three of users are those with offline and engagement, where AI as a or users include individuals with social anxiety or extended use or toward human connection and users, by contrast, social offline and in such AI interaction is more to human such as or human By these distinctions we that not vulnerable users The design challenge in and that emotional AI as a rather than a substitute authentic human Human Emotional If emotional AI is to with human psychological it must be and within an ethical that human relational and emotional agency et al., 2025). This means systems that support rather than human ethical and for AI systems design and 2025). In with our authenticity here is understood as intersubjective mutual responsiveness with rather than as a felt ethical guardrails cannot be they must be attuned to contexts, as assumptions intimacy, care, and agency across What authenticity or boundaries in emotional interaction may culturally in Global South where communal and relational models of are more emotional AI should be its Users must be that they are interacting with a machine et al., 2025). should anthropomorphic or the of emotional reciprocity. Emotional AI should not simulate or therapeutic intimacy without and boundaries 2025). these systems should be designed to and human than solipsism, emotional AI could be to users to engage with emotional or for real-world interactions et al., 2024). than replacing it could users therapeutic between and 2024; Rubin et al., 2024). and must to ensure that emotional AI is not merely but This requires ongoing the psychological of emotional AI, the of and the of and perspectives in the design 2023; et al., 2025). From to To ensure that ethical emotional AI must be designed with that are and outcomes that are each of the three and design This for that are not but testable and psychological risks of dependence and solipsism requires and of the artificial than the illusion of to should and their 2023). extended or emotionally charged to user In or systems should also a of as or where users toward human and 2025). These can be through in dependency, conflict and increases in off-platform social engagement. to mitigate structural risks of commodified intimacy and data extraction, design must data Emotional exchanges should not for 2020). users to or their emotional not only but also affirms are through and user trust that intimacy is not reduced to a et al., 2024). Finally, ethical risks by vulnerable requires outcomes that beyond engagement should measures such as social and off-platform social-contact to interactions or real-world relational 2023). in loneliness, or of should to human users toward or The of these can be not only by individual outcomes but also by reduced of human 2024). these guardrails must be to of within the and are by and to engagement, and data of rather than emotional these dynamics requires that beyond design of affective data for anthropomorphism, pathways when vulnerable users and By ethical guardrails to concrete can not only design but also the structural that emotional AI. In ethical emotional AI design must and data user of affective and By explicitly these guardrails to the three risks we offer a that beyond to that emotional AI and more with human By explicitly these guardrails to the three risks we offer a that beyond to that emotional AI and more with human These with emerging such as on emotion in contexts the AI and and from AI risk and on empathy and 2022; 2025). They We are a emotional in which the capacity to feel and is no exclusively to human Emotional AI to simulated companionship, for emotional especially for the with this the challenge of We must not only what these technologies can but what they are to The of pseudo-intimacy not in its but in its When emotional for relationships, we risk our capacity for empathy, and mutual In over responsiveness over reciprocity, we may emotionally yet The is not to emotional AI, but to it in human We must for and emotional that human intimacy not only but The outcomes discussed or diminished be as hypotheses rather than through These risks can be summarized across three emotional AI fosters pseudo-intimacy and solipsism, guardrails such as and through in and commodified intimacy data and user through and trust vulnerable users which call for and human through off-platform social-contact ratio. we design machines that can mimic not of the and of emotional
Babu et al. (Thu,) studied this question.