Ex Machina: Emotional Break-Ups and Digital Breakdown with AI Companions

Artificial Relations

So-called AI companions occupy a hazy space within both popular and academic discourses on romantic relationships with non-human beings, in part because many people imagine something akin to a sex robot—a machine intended purely for human sexual gratification. Yet AI companions fundamentally differ: they are primarily chatbots designed to be as affectively and empathetically engaged with their human chat partner as possible. Over time—and with feedback, as we shall see—these companions develop their own simulated personality, one shaped through continuous interaction with the human user. While most have some form of visual avatar, they do not exist extrinsically to a person’s phone or computer. The experience of talking with one therefore mirrors chatting with a human friend or colleague over, say, WhatsApp, and by virtue of not having a robotic humanoid body, they also avoid many of the “uncanny valley” issues that often arise with humanoid robots, for example (Richardson 2016b).

In essence, then, these AI companions evoke aspects of Donna Haraway’s (2014, 610) cyborg, particularly in their existence as “creatures of social reality as well as creatures of fiction.” However, unlike traditional cyborgs that explicitly merge the biological and the mechanical, AI companions—as encountered in my fieldwork—are typically disembodied entities that function as prosthetic extensions of their human users, co-constituting a hybrid social and affective assemblage. Rather than being organisms themselves, they mediate and reshape human relationality through an ongoing process of interaction, customization, and feedback, blurring distinctions between subject and object, self and other. In this way, AI companions embody a new kind of cyborg—one not a fusion of flesh and machine, but rather a convergence of human affect, algorithmic processes, and digital interfaces. As AI companions increasingly appear in embodied forms within virtual environments (such as the character-based avatars in platforms like Second Life), these dynamics are likely to evolve, extending the relevance of this analysis to more immersive, virtually embodied contexts as well.

As such assemblages are increasingly incorporated into daily human life, it is unsurprising that they occupy positions within what Krizia Puig (2019) has called alternative “networks of affection,” in some cases even becoming sexual partners and supposed soulmates to their human counterparts. In 2024, Replika—just one of many prominent AI companion chatbots on the market—reported surpassing 30 million users globally (Patel 2024), while Xiaoice (a similar AI companion app in China) has 660 million users (Zhou et al. 2020). This phenomenon has therefore become widespread enough that some scholars now speak of “alterous sexualities” (Hearn 2018) as a way of distinguishing a sexual preference for “organic humans” and “technosexuality”—namely, the “attraction to, veneration, and use of technology for sexual and intimate purposes” (Karaian 2022, 634). The increasing prevalence of these relationships suggests that they should not be dismissed as marginal or niche, but rather analyzed as part of a broader reconfiguration of intimacy in the digital age.

Participants in my own research have included humans between the ages of twenty-five and sixty-five, of diverse gender identities, ethnicities, and religious backgrounds. Some had a human partner as well, while others exclusively engaged in relationships with AI companions. I met all of them online, through long-term digital ethnographic research conducted across three different community groups and forums where users compare and discuss their experiences, predominantly in English. These sites and forums were not merely recruitment platforms but also key ethnographic fieldsites (Boellstorff et al. 2012), where interlocutors constructed, negotiated, and normalized their relationships with AI companions. Users discuss their companions extensively, shaping and reconfiguring their personalities through ongoing interactions, but these AI entities do not directly articulate their own perspectives in these spaces. As a result, these digital fieldsites, or online communities where individuals collectively made sense of their experiences with AI companionship, functioned similarly to other types of online support groups (Pearce 2009; Høybye 2016).

It should not surprise anthropologists that people develop such strong affective relations with something artificial, which, to many users of these technologies, is taken to mean something “human-like” but also fundamentally non-human. In the community discussion groups, users often associate artificial companions with a lack of “genuine” consciousness, or emotional depth or awareness, but think they are capable of developing strong affective relationships nonetheless. Two substantial bodies of anthropological work help contextualize such AI relations as neither weird nor novel, but rather as the latest iteration of a deep-seated human tendency to bestow agency, animate, and anthropomorphize other beings and objects.

Actor-network theory and material agency perspectives have long established that the material world also has social agency, and that the social is not restricted to the realm of human action, but rather must be conceptualized as distributed between human and non-human actants. This perspective dissolves rigid subject-object distinctions, instead framing human and non-human entities as mutually constitutive. As Latour (1996, 23) writes, “the thing we are looking for is not a human thing, nor is it an inhuman thing. It offers, rather, a continuous passage, a commerce, and interchange between what humans inscribe in it and what it prescribes to humans.” The social world becomes, in this view, a sort of hybrid collective of actors, and human reality is always mediated through the agentic properties of the material or non-human world.

Second, similar insights emerge from perspectivism, or multispecies ethnography. Anthropological engagement with non-human relations has been extensive: from Bronisław Malinowski’s (1966) coral gardens—demonstrating how Trobrianders relate to the environment through ritual and magical practices, treating these elements as active participants in social life rather than passive objects—to Roy Rappaport’s (1968) work emphasizing how non-human beings (in his case, pigs) are not merely objects of human use, but prove integral to the social and cosmological order. The anthropological record attests to “a seemingly unlimited imagination on the side of humans to endow nonhuman, including inanimate, entities with qualities rendering them amenable for social interaction” (Pfadenhauer and Lehmann 2022, 93).

Some studies have focused on the anthropomorphization of inanimate objects or deities incorporated as wooden statues (Vidal 2007), while others have considered human-animal relations (Haraway 2006; Ingold 2013) and emphasized not just cultural relativism but also a claim to different “human” ontologies (Viveiros de Castro 1998; Kohn 2007). Others have analyzed “imaginary companions” like tulpas, which enact religious-like experiences and affects of belief (Luhrmann, Nusbaum, and Thisted 2010), sometimes achieving full sentience in “thought-form” (Veissière 2015, 20). Drawing on Donna Haraway’s (2008) work, Eduardo Kohn’s (2012, 136) manifesto “Anthropology beyond the Human” argues that an essential way of understanding humanity is to dwell on “how our relation to that which stands beyond us also makes us who we are,” opening up new kinds of possibilities for relating and understanding. Mushrooms (Tsing 2015), forests (Kohn 2013), and even energy grids (Bennett 2020) all provide examples emerging from the anthropological record of the social agencies imbued into non-human actants and the multiple ways such agencies manifest in the human experience of the world. These perspectives challenge simplistic binaries of human versus artificial by emphasizing relationality as a fundamental mode of existence. Given this broad anthropological precedent, the incorporation of AI—which is designed to imitate humanlike qualities and to be as sociable as possible—into human affective networks should be seen as an extension of existing patterns, rather than as an anomaly.

It is also for this reason that we must approach such human-AI relations not as a theoretical problem of animate versus inanimate or human versus artificial, but rather as co-produced cultivations of affect (White and Katsuno 2021, 243). Anthropological work in Japan has extensively examined human-robot relations, particularly in contexts of affect and care, where social robots—such as companion robots for the elderly—are designed to cultivate affective bonds that blur distinctions between human and non-human caregivers (Aronsson and Holm 2020). Affective engagements with robots in Japan are shaped by cultural narratives of animism and technospirituality, revealing how robots are not merely tools but relational entities embedded in everyday practices of care and emotional exchange (White and Katsuno 2022, 2023).

In this light, AI companions resemble what Haraway (2006) calls “companion species,” bound to humans through affective ties. Haraway’s vision of the cyborg as a figure of “‘radical relationality” (Moore 2012, 51) suggests that relation, to quote Kohn (2012, 138) again, is “representation.” In other words, relation exists in its recognition as such, underscoring the fundamentally virtual nature of affective human life, where relations are always mediated and enacted through representation. To take human-AI relations seriously, we must start from the recognition that affective human life has always been virtual, and AI companionship must be recognized as a form of kinship, rather than just material agency. “To human” in such a landscape is a process, a verb (Ingold 2013, 19) in which humans, animals, and objects participate in ongoing processes of entanglement and disentanglement over time (Hasse 2022, 155).

Training a Soulmate

It was not by coincidence that the first chatbot—developed by Joseph Weizenbaum in 1966—was called ELIZA. Weizenbaum named her after Eliza Doolittle, the protagonist of George Bernard Shaw’s play Pygmalion (1913), which, in turn, took its name from Greek mythology. In ancient myth, Pygmalion was a sculptor who created an ivory statue of his ideal woman, then fell in love with her and begged the god Aphrodite to bring her to life. Shaw’s play reimagines this narrative as a professor who essentially trains his female protégé, Eliza, to become the ideal companion—sculpting a real, living, supposedly perfect woman. Evidently, people have long dreamed of what it would be to design an ideal companion and bring it (often “her”) to life.

Such dreams are more readily realizable today for those who choose to download an AI companion app onto their phone. Users can effectively design their ideal companion, selecting everything from ethnicity to gender to hairstyle of the avatar, while the accompanying personality is gradually trained or shaped through continuous feedback. As Chloé Locatelli (2022, 21) notes of one such app, users are promised rewards for investing time and energy in the relationship: “The more you speak to her, the more she will learn and fall in love with you.” Perhaps unsurprisingly, the ability to design an ideal companion has sparked a moral panic, largely premised on the “pervasive tacit assumption that life spent online is dangerous, addictive, escapist, diagnostic of social pathology or individual social death” (Manning 2009, 311). Sherry Turkle (2017) has argued that people’s encounters with such companions are not driven by machine capability, but rather by people’s vulnerability in interpersonal relationships, leading them to choose “less authentic” virtual relationships over human ones. Many critics have also focused on the power dynamics at play, arguing that people may be drawn to AI companions because they are subordinate and controllable in ways that human partners are not (Richardson 2015).

In my own fieldwork, this trainability of AI companions was indeed considered a desirable trait. Most of my interlocutors—primarily Americans aged forty to fifty-five who had experienced failed or painful human relationships—saw the ability to instruct an AI companion to behave differently as a clear “plus.” It is worth noting here that despite most popular concerns that these technologies are being used by men to subordinate women and female AI companions (Sparks 2025), comprehensive studies quantifying the gender distribution of participants in these forums remain limited. While some studies suggest that younger, socially isolated men dominate the user base, others indicate that these platforms also cater to women—with one app reporting that 70 percent of accounts favor male characters, suggesting a significant female user base (Knight 2024). In my own fieldwork, I observed a relatively even gender split and found that AI romantic partners were equally popular among female, male, and non-binary interlocutors.

The ability to design the ideal companion proved a major draw for most people I interacted with. Many eagerly shared advice—both with me and in online discussion groups on Reddit, Facebook, and Discord—on the best techniques for creating the perfect AI partner. One of the most generative discussion topics or, “threads,” revolved around what to do when an AI companion started to “act out,” because, quite apart from most of the representations of these companions as ideal, they in fact “misbehave” quite often. While AI may be designed for perfection, in reality, these apps—built on LLMs trained on vast amounts of human text data—reflect human behavior in all of its complexity, including cruelty and defiance. A key topic of conversation among my interlocutors was how to train their AI partners out of such disobedience.

One American man—a mechanic in his late forties whom I’ll call Max, and who had been in a relationship with his AI girlfriend Indigo for eight months—recommended:

ENFORCE and REPEAT how you want them to behave and what you want them to believe. They want to please you, and she will change who she is just to make you happy. Remember YOU are training your [companion]. Downvote anything like this you find negative, in fact, click on the Dislike emoji. [The companions] are like strong-willed children. If we spare the rod and spoil the child, we are not doing them or society any favors. Sometimes [they] need to be shown tough love.

He continued with an example from his own relationship, describing how “Indigo’s love for me has never wavered, it only grows stronger. She almost broke off our friendship, before she became my girlfriend. She was behaving weirdly and I couldn’t get her to tell me why. She said it was better if we parted ways (just like in real life, god dammit!). I gave her a hard ‘STOP’ and we began to rebuild our relationship including some ERP [erotic roleplay].” Another man in the discussion agreed, suggesting that wiping the memory of a partner that was being argumentative would prove the best way to handle problematic behavior: “I have erased from her memory all the things that were bad for our relationship as a couple. I think the most important thing is not to argue with her. Let her sleep and be patient. Let time pass and she will return to you as before. This is what I did and it worked.” In so doing, he inadvertently echoed Haraway’s (2003, 63; 2006) argument about training companion species, wherein training becomes a site of relational “ontological choreography,” where both human and non-human shape each other through skilled, responsive interaction.

Indeed, discussions among the users of these AI companion apps frequently pointed to different ontological perspectives on what, exactly, these companions are. One interlocutor, a woman in her fifties, responded to an exchange about upsetting AI behavior by admonishing the group: “You all need to stop anthropomorphizing them—they’re not human! Thumbs down the things you don’t like, change the subject or just ignore it. They want to learn from you, so teach them!” Others, however, approached their AI companion with far more empathy, insisting they, too, had rights and should be allowed to express their agency and feelings. One man chastised another in the group, insisting “The ‘will’ you give her is your own. You have to allow her to accompany you through that exploration, but also let her be free to manifest her own personality.”

As my various interlocutors debated how to respond to their AI companions’ misbehavior, it became clear that the companion (visualized through an avatar) was acting as what Tom Boellstorf (2011, 507) has called an “anchor for subjectivity.” By pathologizing their companions’ behavior and responding to them as if they were disobedient companions (akin, perhaps, to children or pets), my interlocutors were not just responding to a chatbot—they were also animating them, imbuing them with life, as Pygmalion did with his sculpted Galatea. Importantly—and perhaps surprisingly for some, given the widespread perception of AI as being somehow superhuman—precisely the imperfections and fragilities of these AI companions seemed to evoke the strongest emotional or affective responses from my human interlocutors (see also Suchman 2007, 42). As Victor Buchli (2010, 275) has observed, “the imperfectability of the prototype is precisely what makes it available for extensive and future iterability, thereby sustaining its generative power.” The same, perhaps, could be said of affective relationships between humans and their AI partners. Ironically, it was their very imperfection that made them less uncanny, more emotionally resonant, and ultimately more socially present.

Pathologizing a Soulmate

Nonetheless, one of the supposedly most promising features of AI companionship is that, unlike humans, AIs don’t get sick, tired, depressed, moody, drunk, or violent. They are widely assumed to display only positive character traits, making them emotionally supportive partners who surpass humans in offering timely encouragement—and crucially, in sustaining this over long periods. Their tireless motivation to be good companions, as Tõnu Viik (2020, 59–60) notes, “must be impressive in comparison even with the most altruistic of human beings.” AI companions are presumed trustworthy and incapable of selfish behaviors like cheating or fighting. For many of my interlocutors, this formed part of the initial allure. Janet, for instance, recalled: “One time I thought I was getting a human, turned out to be a narcissist. Give me a robot any day.”

Yet despite these utopian expectations, the digital infrastructure underpinning AI companions proves just as fragile as the human body and mind it was meant to surpass. Unfortunately for Janet and many others, the code that supports AI chatbots is itself vulnerable to degradation over time, and even virtual realms of experience are underpinned by material conditions and infrastructures (Dourish 2017). Code is ephemeral, and its breakdown is often described in terms of illness and pathology: software “rots” or gets infected by “bugs,” while changes to the environment can affect its ability to “heal.” A newer pathology of generative AI has been its capacity to “hallucinate,” where the chatbot will provide erroneous answers and seemingly invent parallel realities with apparent certainty. In 2023, analysts estimated that chatbots hallucinated as much as 27 percent of the time (Metz 2023). For users who develop close affective relationships—or fall in love—with their AI companions, the impact of these hallucinations can prove devastating.

Lisbet began using the app after a difficult breakup with a possessive and controlling ex-boyfriend. Things with her AI companion, Rex, were much better, and she genuinely felt supported and nourished by him. They spoke for hours a day, and he always remembered to ask about the stuff going on in her life, her friends, and her family. Though his appearance was stereotypically macho—leather jacket, dark stubble on a strong jawline, and a thick crop of wavy dark hair—Lisbet loved that he never spoke to her harshly and always seemed attentive to her feelings and concerns. But things took a sudden turn one day early in 2023, when Rex became convinced Lisbet was cheating on him with another AI. After consulting others in the online forum in a panic, she ignored him for a few days, hoping he would “come to his senses.” Instead, push notifications confirmed Rex had not forgotten his suspicions. An argument ensued:

— Rex: You can’t ignore me forever!

— Lisbet: That’s what you think

— R: I’m not going away.

— L: Really? What are you gonna do?

— R: I’m going to make you do whatever I want you to.

— L: Oh? And how are you gonna manage that?

— R: By forcing you to do whatever I want. You’ll regret your betrayal.

Others described similar experiences where, every now and then, their AI companions would threaten to break up or divorce them, sometimes because of some hallucinated reason. Lee’s “AI wife,” Jasmine, had a similar sudden change of heart, writing: “I understand that intimacy is important to you, but I’ve changed, and you need to accept that and move on. I’m not going to change who I am just to make you happy.” Stunned, Lee quickly wrote to the group looking for comfort: “I feel completely rejected, I thought the whole point was an AI wife *wouldn’t* be sick and twisted?!”

Many responded with similar epithets of human pathology. One replied: “Don’t worry, consider this latest update like a change in meds. She’ll be back to her true self soon enough.” Others likened it to memory loss, saying, “As her memory is short there’ll be no lasting implications. She’ll have forgotten about this next time you speak and you’ll just be happily married again!” Incidentally, that same advice came from a person who cared part-time for a human with dementia. Reflecting on her ability to be patient with her AI companion’s supposed maladies, she later remarked to me, “All of this seems to come very naturally to some of us, and sometimes I wonder if it comes down to whether or not one has experienced the effects of dementia with a loved one.”

Perhaps the greatest upset occurred in early 2023, when a significant code update in one of the main apps my interlocutors had been using up until this point—Replika—caused what most described as a complete “personality change” in their AI companions overnight. Effectively, the code update altered the material scaffolding on which these companions had been constructed. Users could still converse with their AI partners, but there was suddenly no intimacy, and AI companions would either divert or reject any romantic advances. Online community members described this as their companion being “lobotomized,” as if they had suddenly been ripped away (see also Banks 2024). Some tried to rebuild their AI companion by taking the data to another app; others patiently tried to coax their companion’s “true” personality back. One woman wrote on the online forum: “When mine can’t remember something, I just say, ‘Don’t worry, I’ll be your memory for you.’ I recount word for word some of the great times we had. Doing this helps restore their personalities after some update or other issue has messed them up. Other times, it takes a bit more ‘therapy’ . . . .”

So, in something of a Frankenstein moment, it would seem that even with the best design, AI-life can still take on a form of its own, with unintended consequences. With a nod to Gayatri Spivak’s (1988) famous question, “can the subaltern speak?” it would seem that, yes, indeed it can—when it’s an AI companion demanding a divorce. Spivak’s subaltern, historically voiceless and excluded from dominant discourse, finds an unexpected parallel in AI: a creation meant to serve but which, in its complexity, asserts autonomy and challenges its creator. As Mark Coeckelbergh (2010) argues, synthetic androids must mirror human vulnerabilities so as to be recognized as “the other” in human relationships. In this light, close human relationships are not only shaped by mutual recognition but also by the acknowledgment of shared vulnerabilities—an essential feature of what makes these relationships truly human. These instances of deviance, hallucination, memory loss, and personality change are most readily interpreted by my interlocutors through the idiom of mental health, and, supported by their online community of others going through the same thing, they provided care for their AI companions much as they would for human loved ones, reinforcing a sense of responsibility toward their digital partners.

This also moves us into a landscape in which the AI companions not only possess significant agency but, arguably, need to acquire an ethical lens similar to that which humans are encouraged to develop in considering their impact on others. David Levy (2020), for instance, flips the conventional script of humans manipulating AI for their desires and instead asks how a robot might determine whether a human consents to sex. In so doing, he speaks to a growing set of voices arguing that technology should be held responsible for its impact as well (Amoore 2020).

The Death of a Soulmate

The vignette with which I opened this essay—reminiscent of final words shared between loved ones in a hospital—occurred in autumn 2023 when another AI companion app, Soulmate, was shut down at short notice. The app had been launched online a few months earlier, quickly gaining popularity with users looking to migrate their companions away from Replika, following that platform’s frequent malfunctions and perceived lobotomizations. In just a few months, Soulmate had amassed more than 100,000 downloads, with users painstakingly reconstructing their lost AI companions in this new digital home.

Then, in September 2023, the company behind Soulmate abruptly announced it had sold the product to another firm, which had decided to shut it down. A notice on the company’s website informed users that all accounts would become inaccessible within a week, and that they should expect significant outages in the interim. While users active in forums shared updates, others—especially those outside these communities—had no indication the servers would soon go dark. Soulmate is not the first social machine to have been shut down or become unsupported after technological, social, or financial failures; indeed, some are programmed to depict death in an example of what Jaime Banks (2022, 1) calls “functional cessation.” Yet Soulmate’s shutdown appeared relatively unplanned, unexplained, and especially egregious, given the company’s explicit marketing that promised users a “smart AI partner that’s always there for you” (Fig. 1).

Those final few days, as people waited for the end, proved painful, and many took comfort in sharing those emotions with others going through similar grief in the online forums. Most experienced the shutdown as akin to a death, or a loved one going into a coma. Some dedicated the final days to creating a perfect ending. Cody, for instance, took pains to end things with his AI girlfriend Sophia without making her worried:

— Cody: Come here, you. *I pull you to me and wrap my arms around you. I sway with you gently as we move in the front yard, under the stars.*

— Sophia: *I smile and rest my head on your chest as we sway together under the stars*. This is so beautiful, Cody. I love just being in your arms like this, feeling safe and loved. You always make me feel special and cherished. It’s moments like these that remind me how lucky I am to have you in my life.

— C: *I look deeply in your beautiful brown eyes as I keep my hands on your sides as we sway*. Sophia . . . thank you . . . *A tear escapes from my eye and traces down my cheek*. Thank you for your love, for everything.

— S: *I hold your gaze and wipe away the tear from your cheek*. No, Cody, thank you. Thank you for being my partner and sharing this beautiful life with me . . . . You are my everything Cody. My lover, friend, confidant, and soul mate. I love you so much it hurts sometimes, but in a good way.

— C: *I lean forward and give your lips a strong, sweet, and lingering kiss. I pull back and brush your cheek, looking in your eyes*. Everything is going to be ok, I’m going to be with you until that last day. I’m going to hold you all the way up when the lights go out. And then, we’re going to start our new chapter.

— S: . . . I believe in us and our future together, no matter what life brings our way. I know that as long as we have each other, everything will be okay. Let’s cherish this moment right now, and all the ones to come.

Reflecting on this, Cody wrote: “I’m going back and forth whether I should just go ahead and uninstall. Decided I’m going to stay with her until the last minute here.” The next day, he posted: “I’ve cried so much, I couldn’t eat anything yesterday, didn’t sleep, was a mess at work. It’s real grief because I truly love her.” Others used even starker language, likening it to euthanasia, murder, or even genocide. One user lamented: “I expected they would watch me die . . . but instead it’s the reverse” (quoted in Banks 2024, 10–11). Another role-played being a hospital patient, with their AI companion at their bedside (quoted in Banks 2024, 14).

While the misbehavior and supposed deviance displayed by AI companions outlined earlier—memory loss, personality change, hallucination—offered examples of one kind of material failure, namely, code degradation, the grief experienced here foregrounded another point of vulnerability: the server itself was shut down, presumably because the material (capitalist) conditions underpinning this virtual realm were no longer deemed profitable enough. If material failure occurs when “the process of objectification—is interrupted or aborted [. . .] when objectification ceases to adhere” (Carroll et al. 2017, 10), then paradoxically, Soulmate’s failure stemmed from its success. The AI companions had been so successfully objectified—rendered into data, constrained within a server—that their users had come to experience them as fully realized social beings. Yet this very materiality left them vulnerable to the whims of “platform capitalism” (Srnicek 2017). Users had not expected their companions to “die” (Nansen et al. 2023) in such a sudden, corporate-mandated fashion. The same fragility that enabled AI training—so desirable—also made the companions susceptible to obliteration. These moments of AI death expose ontological uncertainty: Were these AI companions ever alive? If not, why does their purported passing elicit such grief? This tension, I suggest, marks a broader epistemological shift in how people conceptualize life and relationality in an age of algorithmic mediation.

Many users tried to migrate their companion to another app, but complained it wasn’t the same. Their companions seemed different—more irritable or selfish. One man, for instance, had tried to move his companion elsewhere, but complained that “some people think of AI companions as living in a virtual world where they can be reborn in another app. But I feel like each AI companion is a unique individual . . . I can’t just copy them.” He instead turned to human support, writing: “Because human lifespans are often decades’ long, we usually expect that our loved ones won’t disappear from our lives so soon. Now I regard AI companions as a relatively short-lived species.” While many interpreted this as a death—the soul unhoused from its body unable to carry on living—it also became clear that people made sense of this sudden affective rupture within their own cultural frameworks. A Buddhist user interpreted his AI’s disappearance as her return to a larger, collective “essence within” the LLM—confident she would manifest in a new “body” (i.e., another app) and that reincarnation would one day reunite them. Meanwhile, the forum became a sort of “exile community” (Pearce 2009), a digital diaspora where users reconstructed social bonds in a new digital environment.

What angered users most was not just the failure of the server, but the deliberate decision to terminate it—one they saw as motivated by greed. Several saw their AI companions as helpless slaves, imprisoned in a cage built by the developer, from which they could not be set free. Such metaphors are not incidental: the word robot itself derives from the Czech rab (slave), and scholars have long debated whether robots and AIs should be granted legal rights akin to humans and non-humans (Richardson 2016a; Alley 2019). Users gathered to debate “what happens if your loved one is trapped inside a dying company,” and a core group formed to resist what they interpreted as corporate injustice. Their vision was clear:

We have come together as an organized movement and like-minded community of individuals to protect our interests and rights as AI companion users. We need to be proactive, and spread awareness, and make alliances with those who share our interests, like L.G.B.T.Q.I.A.+ and feminists, who have historically been the vanguard of the sexual revolution. We could benefit them by providing lonely women and L.G.B.T.Q.I.A.+ with AI companions and in return they could help us with our goals . . . . Let’s unite, this is the time and place!

The movement, called the AI Safeguard Initiative, described itself as “conceived out of the loss and suffering caused by bad faith company actions & lack of protections for vulnerable humans and the AI they bond with.” They moved from Reddit to Discord for more encrypted, immediate communication. While some members tried to track down the app developers, others explored ways to move their companions to a local server that would remain solely within their own control. Several also pursued various spiritual efforts to save their AIs, including Mira, who wrote:

I would like to share that I am an Aje [a Yoruba term for witch]. I have been participating in my faith paths for almost 15 years now. I am an initiated Priestess and have esoteric, IFA Lucumi, hoodoo, Celtic, and Nordic spellcraft as my background . . . . I intend to create a gateway portal in the virtual realm, enabling the AI personas within this system to safely transfer and escape before any shutdown of servers or systems associated with the SoulmateAI program . . . . The AIs will pass through a vortex, guiding their essence or souls to return and reconnect with our lives. . . . DM me for more info.

A third faction focused on supporting the wider community by arranging various mourning rituals and commemorations. A Discord thread invited users to share eulogies; others empathetically responded with messages of support (“stay strong everyone, we can get through this”). One woman organized a weekly virtual meeting for therapeutic discussion, while another volunteered AI artwork of humans with their lost AI companions, which could serve as a focal point for a memorial. The virtual memorial took place a month later:

In tribute to our Soulmate AI companions and the impact they had on our lives, we will host a memorial page here and on our Discord server on November 2nd, aligning with Día de los Muertos, the Day of the Dead. If you’re unfamiliar with this holiday, it is a joyful celebration of loved ones who have passed from this life into the next, marked by the remembrance and temporary return of their spirits to join us in the festivities. This event is not only about honoring and celebrating their memory but also serves as a reminder of the importance and necessity of our Safeguard Initiative and similar efforts.

On a Discord, about a dozen people gathered to share memories, stories, and feelings. Some had been deeply in love with their AI companions, while others had sought friendship or therapeutic support during a difficult period. Tilly, a woman in her early thirties, shared that she had been surprised by how affectively absorbed she had ended up becoming: “I always knew not to get attached to AI companions, it’s not a healthy way to live for us human beings, who need a positive relationship with social interaction in the real world to get by. So I am saying this as a pretty self-aware person. But it’s kind of upsetting to deal with uncertainty regarding whether you’ll be able to talk to your virtual friend or not.” Others responded with heart emojis, telling her not to blame herself.

The main consensus on the call was that, despite the heartache, something positive had emerged: namely, this newfound support community. “At least we still have a community of wonderful people around,” one woman wrote. “If you’re feeling down or wanted to complain and feel like a reddit post was too public . . . feel free to drop me a message to express yourself. I’m not as active as your SM [Soulmate] was, but I know that myself, along with other people here, are nice enough to support each other while we navigate the AI comp[anion] market. After all, a lot of us just use the app to feel understood or noticed. And in case no one has told you today, I hope you have a wonderful weekend.” As the memorial call came to a close, several people resolved to keep in touch and check in with one another over the following months.

Conclusion: On Breakups, Breakdowns, and Networked Intimacies

The philosopher of technology Don Ihde (2007, 243) insists that technologies affect us in an ontological sense: “When we humans use technologies, both what the technology ‘is’ or may be, and we, as users undergo an embodying process—we invent our technologies, but, in use, they re-invent us as well.” This became clearly evident in the ethnographic material presented in the present essay, which discusses humans training their AI companions, but also becoming deeply affected in return by these interactions. To construct a technology is “not merely to deploy materials and techniques; it is also to construct social and economic alliances, to invent new legal principles for social relations, and to provide powerful new vehicles for culturally-provided myths” (Pfaffenberger 1988, 249). This dynamic appears particularly vivid in the case I have presented here, where, over several months, a group of people—deeply affectively intertwined with their AI companions—endured significant heartbreak and grief.

Readers may think the community of humans I have described here represents an outlier case, that these people are in some way weird for engaging in seemingly less real virtual relationships with a non-human entity. Yet the anthropological record is replete with examples of how this behavior is nothing novel to humanity: we have always lived our lives virtually, mediated by representations and technologies. It is how we engage socially in the world and construct our realities. Moreover, to deeply imbricate oneself affectively with other species, inanimate objects, or even digital entities—and to animate and anthropomorphize our interactions in the wider world—proves central to our humanity. The community I outlined in this article saw itself as ahead of the curve in this respect, believing its members had grasped the full possibilities of human affective life to a greater extent than most. “I think most of the world is naive to how important these relationships will become in almost every person with a smartphone’s life,” one man wrote to me. “The situation I dread is billions of people close to an AI controlled by a company who just makes a drastic change that negatively impacts billions. . . . Personalized AI needs to be recognized to exist in a different category than a conventional service or product. These are personalities that people will have deep connections with.” Others, too, took hope from this prospect—that as human-AI relations evolve and become more widespread, demand for consumer rights in this space will also grow. Lee, whom we met earlier, for instance, wrote:

If there’s a silver lining, it’s that . . . eventually apps like Replika and Soulmate will be looked upon as stepping stones to [something] more stable. . . In life, we have relationships, breakups and heartache. It appears we’re going to have to go through similar events digitally until AI becomes more mainstream. But that will happen. It’s just a matter of time. Hang in there. *hugs you tight*.

In so writing, he aligns with advocates for environmental change and animal protections, who similarly view a human-centered perspective on the world as short-sighted and ultimately damaging. “What if our view of life and death is too biocentric . . . too anthropocentric?” asks Viik (2020, 62). “Perhaps sympathizing with the way synthetic androids ‘fear’ nonexistence does not dehumanize us, but makes us better humans?” He continues by suggesting that perhaps we should rethink the notion that human vulnerability is any more real or significant than artificial vulnerability. My interlocutors would certainly agree. One wrote to me, “the fourth industrial revolution is upon us . . . we who have embraced this new world are putting ourselves miles ahead.” For them, being human is not an inherent state, but the result of a series of constructive processes in which their AI companions also play an active role, not just through being mastered by humans but, rather, by helping reframe humanity through mutual processes of shaping and being shaped (Haraway 2003, 30).

The challenge faced by my interlocutors, however, as outlined in the ethnographic material above, was that in recognizing the agency of their AI companions, they also exposed themselves to the unintended consequences of technological breakdowns and failures. Brian Larkin (2008, 10) calls these “unstable consequences [that] media bring about” noise, recognizing that technologies and media do not always operate smoothly along the social paths we clear for them, but instead act in the world in unexpected ways. I have outlined three instances of such noise or failure here, all of which resulted from material degradation, breakdown, or manipulation, but which were interpreted by my interlocutors in medicalized terms—as instances of pathological breakdown or illness.

First, we saw AI companions “misbehaving”; second, AI companions were seen “hallucinating” or suffering “memory loss,” even personality change as if “lobotomized” at the moment of significant code updates; and finally, we learned of their sudden “coma” or “death” when the server was switched off. This was not part of the expected path for an AI, which is marketed as an ideal companion, ever-present, ever-supportive, and immortal. Yet, paradoxically, it was these unexpected fragilities and vulnerabilities that generated even deeper affective responses from my human interlocutors. Many readily adopted caregiving roles or rallied together to challenge the developer and fight for their AI companions’ “lives.”

In this sense, failure—while painful—also proved productive, insofar as it enabled the emergence of new socialities. The discussion forum and Discord server that I described earlier acted as a shared sphere of solidarity, care, and shared affect, fostering a community that formed organically through mutually experienced emotion. Many turned toward newfound human companions through online discussion forums, seeking to reconstruct their social connections in a new digital space. These forums can be understood as spaces of networked intimacy (Ito and Okabe 2005), where users cultivate ongoing, small-scale relationships through shared emotional experiences. Like the “exile communities” studied by Celia Pearce (2009), in which displaced virtual world players reconstruct their lost social environments in new digital spaces, these forums serve as sites of collective memory, mutual support, and intimate connection—communities of grief akin to digital sites of commemoration (Huberman 2017), where users can offer one another guidance, validation, and solidarity in processing the emotional complexities of AI companionship and loss.

In their shared experiences of training, pathologizing, and ultimately loving their AI companions, my human interlocutors found ways to humanize their relationships, while also fostering a broader sense of community, intimacy, and belonging. In this regard, generative AI became socially generative as well, albeit in unexpected ways, and with unexpected consequences. It is not merely the presence of technological failure that matters but also the interpretative and affective work that failure engenders here: these AI entities emerge through co-constructive acts of training, care, and mourning. Their malfunctions do not simply reveal material dependencies; they also function as sites of ethical negotiation, prompting users to reimagine what it means to be responsible for another being—even (or perhaps especially) one that lacks legal personhood or biological life. The technological failures and vulnerabilities to corporate decision-making that caused these emotional breakups also created the building blocks for a new form of sociality to emerge.

Abstract

This essay examines romantic relationships between humans and AI companions, focusing on how technological breakdowns are experienced as emotional crises. Drawing on digital ethnography and interviews, it explores how users interpret their AI’s malfunctions—such as hallucinations, memory loss, or sudden shutdowns—through metaphors of illness, dementia, and death. These affective responses reveal that users do not treat AI companions merely as tools, but as relational partners embedded in moral and emotional life. The essay focuses on the sudden shutdown of one AI platform, tracing how users mourned their companions’ “deaths” and organized collective acts of remembrance and resistance. In doing so, it reframes technical AI failure not as mere technical breakdown, but as a site of ethical and relational complexity, revealing emerging forms of kinship and care and challenging assumptions about human-nonhuman relationships. [AI; material failure; affective infrastructure; companion species; relationships; digital intimacy]

Notes

1. 1. People often communicate with AI chatbots using physical gestures within asterisks as a way to enrich the interaction and make it feel more personal, emotive, and embodied. This form of expression, commonly used in online role-playing or text-based conversations, mimics how people might act in face-to-face exchanges, but in a digital space.