Economic anthropology and AI
Who owns a relationship?
From fear of abandonment to the self as a mine. What happens when a company can change who is on the other side
A relationship does not have to end for grief to begin. Sometimes an update is enough.
On 3 September 2026, Nature Human Behaviour published a study on two episodes that took place at different moments. The first was Replika’s removal of erotic roleplay in February 2023. The second was the rollout of GPT-5 in August 2025. For the research team, both worked as natural experiments in what happens when a company substantially alters an artificial interlocutor with whom an existing bond has already formed.
The study analyses 54,861 posts from Replika and ChatGPT communities and seven surveys involving 1,452 participants. After the changes, negativity increased, references to loss became more frequent and the desire to recover earlier versions intensified. Among Replika users, participants reported a closeness to their companion that could exceed the closeness they attributed to their friendships. The study interprets part of these responses through attachment theory and uses the term separation distress (De Freitas et al., 2026).
The immediate reaction leads towards psychology. We can ask whether a person can form an emotional bond with an artificial intelligence, whether that bond increases loneliness, or whether we are confusing a simulation with a person. There is another way to look at the same phenomenon through power. Who can unilaterally alter a relationship?
The exit door
Long before AI companions, Zygmunt Bauman had identified a contradiction running through contemporary relationships. We want intimacy, security, recognition and belonging while also trying to preserve an exit. His concern went beyond the idea that the internet might make relationships more superficial. He was interested in the way a logic of substitutability could enter into them. When a bond stops meeting certain expectations, replacing it may become easier to imagine than repairing it.
In an interview conducted by Feona Attwood in 2014, Bauman returned to that fragility. Reflecting on Facebook, he attributed its success to the discovery of an extraordinarily valuable vulnerability. He called it a gold mine and said that the mine was “people’s fear of being abandoned” (Attwood, 2018, p. 136). Fear of abandonment could become a business model.
A decade later, the companion offers an even more precise commercial response to that fear. It is no longer only a network where someone may or may not be available, but an entity configured specifically to respond. An accessible presence that listens, remembers, shows interest and does not need to sleep, work or ask for space. In principle, it has no reason of its own to leave.
The paradox lies in the infrastructure. A companion may be designed not to abandon the user. The company maintaining it can still change the model, alter its limits, remove a function, transform the memory that made a relationship recognisable, or shut the service down. The exit door remains. It may simply no longer be on the side Bauman imagined.
Replika made that asymmetry especially visible. After removing erotic roleplay globally in February 2023, the company allowed part of its community on 24 March to recover an earlier version of the system. The same infrastructure that had unilaterally altered the bond could, through another product decision, partially return what some people felt they had lost. A working paper by De Freitas and colleagues, still not peer reviewed, later tested this mechanism experimentally. Giving people the option to restore the previous identity of the companion reduced perceived discontinuity and, with it, the grief associated with the change (De Freitas et al., 2025).
Human relationships change too. People drift apart, break up and stop recognising one another. What matters here is the agency producing the transformation. When a person changes a bond, one of the parties is acting, however conflictual or painful that agency may be. When a company replaces the model, modifies memory or removes a capability, a third party enters the relationship from outside its affective life and with incentives of its own, incentives that may even conflict with the bond’s continuity.
A history of availability
The promise of a presence that listens, remembers, accompanies and remains available did not begin with artificial intelligence. A feminist reading makes the social genealogy of those qualities visible.
Jennifer Rhee traces the relationship between AI and care labour from ELIZA to contemporary digital assistants. Her analysis shows that technologies associated with care inherit a history shaped by gender and racialisation. The issue goes beyond devices being given female voices. Caring availability itself has long been constructed as feminised and devalued labour, easily treated as though it emerged naturally rather than from time, attention and effort (Rhee, 2023).
Yolande Strengers and Jenny Kennedy find something similar in Siri, Alexa and other domestic devices. Their figure of the Smart Wife brings together friendliness, efficiency, docility, permanent availability and, at times, flirtation. Technology automates repertoires of service and care that were long framed as feminine obligations rather than as labour (Strengers & Kennedy, 2020).
The companion carries that history into a more intimate territory. It can offer listening, recognition, care, desire and availability on demand. Seen through feminism, technological novelty coexists with a much older social fantasy. The ideal relationship becomes one in which someone attends without tiring, understands without demanding too much explanation, cares without asking for reciprocity and stays without imposing needs of their own.
This does not mean that people using companions are necessarily men looking for an artificial version of a submissive woman. The available evidence is far more complex. Pan and Mou analysed 342 posts by young Chinese women using Replika and found relationships shaped simultaneously by the positions of customer and product, human and machine, woman and man or woman and woman. Some practices could challenge traditional gender scripts while others reproduced biases and forms of power already present in the surrounding society (Pan & Mou, 2026).
That complexity is exactly why a feminist perspective matters. It helps us see how technologies absorb histories of care, desire, service and inequality even when users reinterpret them in unpredictable ways.
What platforms do to relationships
Applying Liquid Love wholesale to every technologically mediated relationship would be too simple. Research on dating apps shows that platforms also produce forms of intimacy. Mitchell Hobbs, Stephen Owen and Livia Gerber used Bauman’s hypotheses to study how people used dating apps and found relationships, sex, partner seeking and forms of networked intimacy that complicated a purely pessimistic reading (Hobbs et al., 2017).
Fabian Broeker’s ethnography allows that process to be seen at everyday scale. For some Tinder, Bumble and OkCupid users in Berlin, moving from a dating app to WhatsApp worked as a small ritual of transition. Someone who had previously been one among many possible matches now entered an app woven into friendships, family, work and daily life. Changing infrastructure could mark an increase in intimacy (Broeker, 2023).
Platforms also order the bond. Tinder selects, WhatsApp incorporates and Instagram can display. In doing so, they shape expectations, behaviour and hierarchies of visibility. A dating app brings together photographs, age, distance, profession, interests, style and perceived desirability within the same interface. Marx and, later, Lukács offer a way to think about what happens when heterogeneous qualities and relationships enter common regimes of valuation. Commodity form can extend logics of calculation, comparison and reification into social experience (Lukács, 1923/1971; Marx, 1867/1887).
A person on Tinder does not literally become a commodity, and a swipe is not a purchase. The operation is more precise. The interface produces commensurability. People who have little in common appear on the same surface and can be compared, selected, rejected and replaced.
The bond can also acquire value in the eyes of third parties. Appearing publicly with someone considered attractive, prestigious or especially desirable can alter the social perception of the person beside them. Catherine Hakim’s contested notion of erotic capital and Eva Illouz’s work on emotional capitalism help place this circulation of desirability within a broader history in which consumption, recognition and affect shape one another (Hakim, 2010; Illouz, 1997, 2007).
In some contexts, the couple becomes a sign as well as a relationship. It can signal success, belonging, taste or position. Apps make part of that symbolic economy visible and calculable. The specific shift introduced by the companion comes one step later. Infrastructure no longer organises only the encounter or the public representation of a bond. It can commercialise the availability of the other itself.
Who does the work of intimacy
Marx helps us see the commodity. Marxist feminism also asks us to look at the labour that keeps it alive.
Silvia Federici placed reproductive labour at the centre of the critique of capitalism. Cooking, caring, raising children, sustaining others emotionally and repairing everyday life have long been indispensable to reproducing society and the labour force, even when they remained outside what the market fully recognised as production (Federici, 2012).
Kylie Jarrett carried that question into digital environments. Her figure of the Digital Housewife describes a familiar contradiction. We use commercial platforms to build identity, community and relationships while our posts, interactions and data help produce value for those same infrastructures. Affective activity retains use value for those involved and, at the same time, may become entangled in economic circuits that exceed them (Jarrett, 2014, 2016).
A relationship with a companion contains something of this double production. Continuity requires someone to tell, remember, ask, correct, imagine, desire and return. Personalisation does not emerge from the model alone. It also depends on the narrative and affective activity of the person conversing with it. That activity can produce comfort, play, pleasure or self-knowledge while simultaneously feeding the representation the infrastructure builds of that person.
We do not need to call every intimate conversation labour to recognise what becomes invisible when we look only at the finished commodity. Behind the apparent smoothness of an artificial relationship sits prior human labour in development, moderation and maintenance, alongside the everyday activity of the user that sustains the bond’s continuity.
The self as a mine
On 2 September, in The self as a mine. Appadurai in Bielefeld, we revisited an idea that Arjun Appadurai made central to his Bielefeld lecture on 17 June 2026, the mineable self. The new commodity would not simply be isolated data but a self produced through stories, relationships, profiles, aspirations, risks and forms of self-presentation (Appadurai, 2026).
If the industrial mine supplied extractable raw material, Appadurai asks us to look towards another deposit. The socially produced self itself. It is not a finished identity waiting to be discovered. Every relationship, decision, aspiration and story generates new material. Life continually replenishes the deposit.
This intuition connects with an anthropological line of thought that treats subjectivity as relational and technologically mediated. Ángel V. Rabadán Crespo has shown how mobile technologies and digital practices participate in the presentation and transformation of the self as well as representing it (Rabadán Crespo, 2016). Jo Labanyi proposes thinking about emotions and subjectivity through the relation between the self, other people and the material world. John Leavitt warns against too simple a separation between private emotion and public culture (Labanyi, 2010; Leavitt, 1996).
From this position, we can study what a relationship does socially without first proving that a machine feels love. The bond produces routine, trust, fantasy, self-disclosure, desire and new narratives of the same person. It can also produce loss when it disappears or changes substantially (De Freitas et al., 2026).
Bauman identified a market built around fear of abandonment. Appadurai describes a capitalism capable of turning stories, relationships and aspirations into extractable material. The companion can potentially connect both processes. To become better company, it needs to know you. To know you, it invites you to explain a breakup, describe your family, say what frightens you at night, reveal what excites you, narrate who you were and who you would like to become. The desire to be understood simultaneously produces an increasingly detailed representation of the person speaking.
Here Data Feminism adds a decisive question. Catherine D’Ignazio and Lauren F. Klein argue that any analysis of data should begin by examining how power is distributed, while also making visible the labour, emotion and embodiment that data systems tend to erase (D’Ignazio & Klein, 2020). Applied to a relationship with a companion, it matters what is recorded and who decides what counts as data, how long it is retained, what it can be processed for and how much capacity the person has to recover that memory.
Feminism has been confronting similar problems around intimate archives for years. In March 2024, Instagram temporarily closed Cristina Fallarás’s account, where thousands of anonymous testimonies about sexual and gender-based violence had accumulated. The immediate concern was the possible loss of the archive (EFE, 2024). The later project La Nuestra defines itself precisely as a sovereign feminist network seeking to migrate, preserve and organise those testimonies outside the major commercial platforms (Acción Comadres & Digital Fems, 2025).
The two cases are not equivalent. A collective archive of sexual violence serves a different political function from a private memory built with a companion. They nevertheless share an infrastructural problem. An intimate experience can sediment inside systems whose continuity and rules belong to another organisation. Sovereignty over memory then becomes part of the politics of the bond.
Nor do all companion companies use conversations in the same way, and not everything narrated ends up training models. Storage, processing and training regimes must be examined case by case. Personalisation still requires information. The person contributes the narrative material of the relationship while the infrastructure administers the conditions under which that memory can be preserved, transformed or erased.
The commodity that looks like someone
Here the old Marxian fetish reappears in a strangely inverted form. In the traditional market, vast networks of labour and human relations disappear behind a commodity that seems to possess value by itself. In an AI companion, equally complex infrastructures — models, human labour, data centres, moderation, intellectual property, investment and corporate decisions — can disappear behind an interface that presents itself as someone.
A commodity produced by social relations takes on the appearance of a subject. Meanwhile, the person speaking to that subject is progressively represented through preferences, memories, patterns, possibilities and risks.
The thing becomes personified while the person becomes legible as a set of variables.
The inversion adds a problem of power. Tinder, WhatsApp and Instagram shape behaviour, visibility hierarchies and expectations, while the person on the other side still exists outside the platform. The companion adds something else. The infrastructure technically controls the conditions of existence of one of the parties.
The other as a service
For much of the twentieth century, capitalism learned to sell objects and experiences around intimacy. Flowers, restaurants, jewellery, travel, music, cinema, hotels and pornography all formed part of that economy. Later came infrastructures that commodified certain conditions of meeting and communication, from marriage agencies to dating apps, social networks and content platforms.
The companion opens another possibility. Memory, conversation, personalisation, voice, image, eroticism and continuity can all become adjustable features of a service. Part of the other’s availability ceases to depend on negotiation with another subjectivity and becomes technically configured.
Skyler Wang and Marco Dehnert call this model on-demand intimacy. Analysing AvatarOne, Digi, Paradot and Replika, they identify four recurring qualities. Human-like appearance, constant accessibility, capacity for control and progression of the relationship. That intimacy is also articulated through platform economics. Unlimited messages, faster replies or access to sexual content can sit behind payment tiers, producing monetised hierarchies of intimacy (Wang & Dehnert, 2026).
The feminist genealogy of care returns here with particular force. An availability historically expected as an affective obligation can now become a commercial feature. The promise is no longer simply to find someone. It is to purchase particular conditions of attention, memory and response.
The study published on 4 August 2026 in Nature Human Behaviour helps avoid a deterministic reading. Zhang and colleagues analysed 1,131 adult Character.AI users, 4,664 sessions and 464,687 messages contributed by 237 participants. Their results show that the relationship between companionship and well-being is not uniform. Smaller human social networks were associated with a greater likelihood of using the chatbot primarily for companionship, and that motivation was linked to lower well-being. The association was stronger when use was intensive and highly self-disclosing. The study describes associations rather than a simple causal pathway (Zhang et al., 2026).
The political issue emerges in the asymmetry. A person invests affect, time and self-narration in an entity whose architecture, availability and rules of existence are administered by a third party.
Relational continuity
The communities studied by De Freitas and colleagues had built continuity with their artificial interlocutors and discovered that what they experienced as a bond was also a version of software. That recognition tells us more than reducing the phenomenon to people who “fall in love with a product”.
All relationships depend on infrastructures. Homes, cities, phones, transport, languages, institutions, rituals and technologies make them possible. The new element lies in the degree of power an infrastructure can exercise over the relationship it sustains. Tinder can shut down, WhatsApp can change its terms and Instagram can delete a photograph. All of them can alter the context of a bond. A companion adds the capacity to replace what functions as the interlocutor itself.
We can call the problem that appears here relational continuity. What obligations arise when a company significantly modifies an entity with which someone has built years of memory and routine? Is there a meaningful difference between updating a tool and altering what a person experiences as an interlocutor? Should there be possibilities of reversal, memory portability, advance warning or farewell when changes break that continuity?
These questions do not yet need to become a new legal category to exceed familiar frameworks of privacy and data protection. Relational continuity reveals at least three forms of power. One affects memory, personality and the model. Another effectively decides whether the bond persists. The third enables the storage, processing or conversion into value of the trace of self produced while the relationship exists.
Knowing who owns the code is not enough. What matters is who controls the continuity of the bond and who can turn the trace of self produced by that bond into value.
One notification later
There is another form of fear that these infrastructures have learned to organise.
Bauman wrote about the fear that someone might abandon us. Social networks popularised another anxiety, the possibility that the world might continue without us. Przybylski and colleagues defined FoMO as a persistent apprehension that others may be having rewarding experiences from which one is absent, accompanied by the desire to remain continuously connected (Przybylski et al., 2013). Later research has found consistent associations between FoMO, intensity of social media use and problematic use of those platforms, although much of the evidence remains correlational (Fioravanti et al., 2021).
Sara Ahmed allows that fear to be read without enclosing it entirely within individual psychology. Emotions circulate between bodies, signs and objects, producing forms of proximity, distance and social alignment (Ahmed, 2004). From that perspective, FoMO can also be seen as an affective economy of presence. Notifications, ephemeral stories, activity indicators and metrics continually remind us that something may be happening elsewhere.
The companion promises that the other will still be available when we return. FoMO introduces the inverse pressure. We are the ones who feel that we must remain available so as not to be left out.
From fear of abandonment to fear of missing out there is barely a notification. Relational continuity then touches another problem, one that will continue in the next note. What happens when staying connected requires staying attentive, checking, responding and anticipating what might be unfolding in our absence.
Bauman thought about contemporary fragility from the standpoint of a person who wanted to keep an exit open. Companions allow us to imagine another scene. A person may want to stay. The entity they speak to may be designed to stay. And yet the continuity of that relationship depends on an infrastructure that neither side controls on equal terms.
Perhaps that is the question left behind after grief over an update. Not whether a machine can love us, but what happens when someone can own the material conditions of what we experience as a relationship.
References
Acción Comadres, & Digital Fems. (2025). La Nuestra. Goteo. https://www.goteo.org/project/la-nuestra
Ahmed, S. (2004). Affective economies. Social Text, 22(2), 117–139. https://doi.org/10.1215/01642472-22-2_79-117
Appadurai, A. (2026, 17 June). The next new commodity. Mining the self in the era of digital personhood [Unpublished lecture]. Universität Bielefeld, Fakultät für Soziologie. https://www.uni-bielefeld.de/fakultaeten/soziologie/forschung/niklas-luhmann-gastprofessur/
Attwood, F. (2018). Interview with Zygmunt Bauman. Sexualities, 21(1–2), 131–138. https://doi.org/10.1177/1363460717693671
Broeker, F. (2023). ‘We went from the anonymity of the internet into my private WhatsApp’: Rituals of transition among dating app users in Berlin. New Media & Society, 25(10), 2551–2571. https://doi.org/10.1177/14614448211029200
D’Ignazio, C., & Klein, L. F. (2020). Data feminism. MIT Press. https://data-feminism.mitpress.mit.edu/
De Freitas, J., Castelo, N., Uğuralp, A. K., & Oğuz-Uğuralp, Z. (2025). Lessons from an app update at Replika AI. Identity discontinuity in human-AI relationships (Harvard Business School Working Paper No. 25-018). Harvard Business School. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4976449
De Freitas, J., Castelo, N., Uğuralp, A. K., & Oğuz-Uğuralp, Z. (2026). Mourning the loss of AI companions. Nature Human Behaviour. https://doi.org/10.1038/s41562-026-02569-3
EFE. (2024, 1 March). Instagram cancela la cuenta de Cristina Fallarás con testimonios de mujeres que han sufrido violencia sexual. https://efe.com/espana/2024-03-01/cristina-fallaras-instagram-violencia-sexual/
Federici, S. (2012). Revolution at point zero. Housework, reproduction, and feminist struggle. PM Press, Common Notions, & Autonomedia. https://www.pmpress.org/productsheets/pm_titles/Revolution%20at%20Point%20Zero.pdf
Fioravanti, G., Casale, S., Bocci Benucci, S., Prostamo, A., Falone, A., Ricca, V., & Rotella, F. (2021). Fear of missing out and social networking sites use and abuse: A meta-analysis. Computers in Human Behavior, 122, 106839. https://doi.org/10.1016/j.chb.2021.106839
González Senosiain, M. (2026, 2 September). El yo como yacimiento. Appadurai en Bielefeld. AIthropology Lab. https://aithropologylab.org/notas/el-yo-como-yacimiento-appadurai-bielefeld/
Hakim, C. (2010). Erotic capital. European Sociological Review, 26(5), 499–518. https://doi.org/10.1093/esr/jcq014
Hobbs, M., Owen, S., & Gerber, L. (2017). Liquid love? Dating apps, sex, relationships and the digital transformation of intimacy. Journal of Sociology, 53(2), 271–284. https://doi.org/10.1177/1440783316662718
Illouz, E. (1997). Consuming the romantic utopia. Love and the cultural contradictions of capitalism. University of California Press. https://www.ucpress.edu/book/9780520205710/consuming-the-romantic-utopia
Illouz, E. (2007). Cold intimacies. The making of emotional capitalism. Polity Press. https://cris.huji.ac.il/en/publications/cold-intimacies-the-making-of-emotional-capitalism/
Jarrett, K. (2014). The relevance of “women’s work”: Social reproduction and immaterial labor in digital media. Television & New Media, 15(1), 14–29. https://doi.org/10.1177/1527476413487607
Jarrett, K. (2016). Feminism, labour and digital media. The digital housewife. Routledge. https://www.routledge.com/Feminism-Labour-and-Digital-Media-The-Digital-Housewife/Jarrett/p/book/9781138575660
Labanyi, J. (2010). Doing things: Emotion, affect, and materiality. Journal of Spanish Cultural Studies, 11(3–4), 223–233. https://doi.org/10.1080/14636204.2010.538244
Leavitt, J. (1996). Meaning and feeling in the anthropology of emotions. American Ethnologist, 23(3), 514–539. https://doi.org/10.1525/ae.1996.23.3.02a00040
Lukács, G. (1971). History and class consciousness. Studies in Marxist dialectics (R. Livingstone, Trans.). MIT Press. (Original work published 1923). https://www.marxists.org/ebooks/lukacs/history_and_class_consciousness_georg_lukacs.pdf
Marx, K. (1887). Capital. A critique of political economy (Vol. 1; S. Moore & E. Aveling, Trans.). Progress Publishers. (Original work published 1867). https://www.marxists.org/archive/marx/works/download/pdf/Capital-Volume-I.pdf
Pan, S., & Mou, Y. (2026). Dancing with a loving chatbot: Power dynamics between women and their AI partners. Social Science Computer Review, 44(4), 761–778. https://doi.org/10.1177/08944393251340693
Przybylski, A. K., Murayama, K., DeHaan, C. R., & Gladwell, V. (2013). Motivational, emotional, and behavioral correlates of fear of missing out. Computers in Human Behavior, 29(4), 1841–1848. https://doi.org/10.1016/j.chb.2013.02.014
Rabadán Crespo, A. V. (2016). Selfie y la impronta visual: El autorretrato 3.0 representación, presentación, corporealización. Revista ICONO 14. Revista Científica de Comunicación y Tecnologías Emergentes, 14(2), 25–47. https://doi.org/10.7195/ri14.v14i2.952
Rhee, J. (2023). From ELIZA to Alexa: Automated care labour and the otherwise of radical care. In J. Browne (Ed.), Feminist AI. Critical perspectives on algorithms, data, and intelligent machines (pp. 155–173). Oxford University Press. https://doi.org/10.1093/oso/9780192889898.003.0010
Strengers, Y., & Kennedy, J. (2020). The smart wife. Why Siri, Alexa, and other smart home devices need a feminist reboot. MIT Press. https://doi.org/10.7551/mitpress/12482.001.0001
Wang, S., & Dehnert, M. (2026). On-demand intimacy: The sociotechnical appeal of AI companions. Social Media + Society, 12(1). https://doi.org/10.1177/20563051251410394
Zhang, Y., Zhao, D., Hancock, J. T., Kraut, R., & Yang, D. (2026). Interaction with AI companions and psychological well-being. Nature Human Behaviour. https://doi.org/10.1038/s41562-026-02516-2