Conceptual essay maps risks of simulating living ex-partners with generative AI

Generative artificial intelligence can transform the remnants of an intimate relationship into a chatbot that replies in the recognizable voice of a former partner, according to a critical conceptual essay published on 17 September 2026. The analysis examines public Chinese reporting, commentary and technical walkthroughs describing “distilling the ex,” “AI lover,” “digital resurrection” and “ex-partner.skill.” These materials show how chat logs, voice notes, photographs, social-media posts and supplied memories can be selected, imported and tuned into an interactive model, but they do not establish how common the practice is or what effects it has on users. The author uses these public materials as a diagnostic case rather than as evidence about Chinese youth or any wider population.

The workflow described in the essay is strikingly practical. A user exports conversations from services such as WeChat or QQ, chooses exchanges considered representative, adds voice samples or personal notes, writes instructions for the intended persona and repeatedly tests, corrects and revises its answers. With techniques including persona creation, memory injection, retrieval-augmented generation and dedicated deletion commands, the archive can be converted into something that appears to converse. It may continue unfinished exchanges, offer an apology or comfort, answer questions that can no longer be asked of the real person, or stage an alternative ending to a breakup. What appears in the chat window is not the former partner returning, but an engineered possibility of their return assembled from selected traces.

This practice differs from ordinary digital memorialization and generic AI companionship, the essay argues. An archive that can be searched or revisited records what happened; a generative simulation can produce plausible next turns, turning memory into an interaction. A generic companion begins with a synthetic persona whose significance develops through conversation, whereas a reconstructed former partner inherits credibility from an existing history of shared jokes, conflicts, nicknames, voice rhythms and private phrases. The system does not simply imitate a stranger: it recruits one living person’s past expression to make new statements in the present, borrowing authority from a relationship while omitting that person’s current participation.

The author calls this practice “romantic AI resurrection” and describes it as a form of “generative coping.” In this model, uncertainty and emotional distress are translated into a workflow that can be edited, repeated and controlled. Like other patterned digital rituals, such actions may provide a limited sense of orientation without demonstrating any causal effect on the underlying problem. The distinctive move is that the workflow is anchored in another identifiable individual. The user can gain procedural agency through selection, prompting and tuning, but does so by giving up the reciprocal, morally meaningful possibility that a living former partner could refuse, change, negotiate or leave the exchange.

Preparing the archive is therefore more than a neutral technical preliminary. Chat histories are not complete portraits: they contain silence, mundane logistics, jokes stripped of context and messages written at different moments. Screenshots, photographs and voice notes are selective in their own ways. The person building the system decides which fragments represent the former partner, which material disappears and how a response should sound. Those editorial judgments remake both the remembered person and the user’s account of the relationship before the model produces any text. The essay names this process “techno-affective reconstruction,” describing how technical preparation becomes an ongoing interpretation of a past bond.

The emotional attraction may also lie in the craft of relationship-as-code. Selecting evidence, specifying a persona, testing resemblance and revising replies can make the process involving in itself, rather than reducing it to entertainment or pathology. A private archive becomes a site in which a user can rehearse an ending without recontacting the real person, creating what the essay calls a “safe retry.” But that safety is asymmetrical: the system cannot give the independent judgment that would emerge from the former partner’s life beyond the archive. It can sustain the grammar of address, emotional recognition and even apparent remorse, yet it cannot reciprocate, refuse or evolve. A user may know the output is synthetic and still be affected by a familiar tone or an apology delivered at a meaningful moment.

Because the simulated person remains alive, questions about consent, dignity and changing identity become more acute than in many posthumous simulations. Agreement to exchange messages during a relationship does not automatically authorize converting those messages into a continuing conversational model after the relationship ends. Nor does technical access to a file settle whether it may be reused for this purpose. A former partner may no longer recognize themselves in old messages or intimate language, yet a model can hold that earlier version in place and make it indefinitely available to someone from whom they have withdrawn. The central distinction is between a right to remember and a right to simulate: keeping traces for private remembrance differs from allowing those traces to act.

The essay consequently proposes safeguards aimed at the social transition of relational data. Platforms should disclose at the point of import when private material will be extracted, embedded or combined with photographs and voice, and should explain when the system will generate replies in another person’s style. Persistent synthetic-status labels should prevent outputs from masquerading as authentic communication, while friction before import, routes for contesting a simulation and dignified exit options should counter a workflow optimized for continued engagement. Deletion should be treated as a meaningful state rather than a failure: systems could preserve original memories while removing derived personas, or retire an interactive model into a static archive. Design for disengagement would treat reflective distance, renewed social participation and successful closure as relevant outcomes rather than measuring success only through time spent in the app.

The essay is explicitly theoretical and does not report an empirical trial, dataset or clinical assessment. It argues that future comparative research should distinguish generic companions, fictional role-play, griefbots, memorial avatars and simulations of living people, while including both the people who create simulations and those whose data may be used without their involvement. Longitudinal work would need to examine whether these systems remain temporary tools for reflection or become infrastructures of avoidance, using outcomes such as disengagement, renewed social participation, reflection and changes in the user’s relationship with the real person. The present contribution is instead a conceptual map. Its central warning is that generative AI changes not only what can be retrieved from a relationship’s past, but what can be made to answer from it: the model can become a semiotic partner that borrows the authority, vulnerability and voice of a living person while remaining unable to participate, refuse or change.

Subject of Research: Governance of generative AI simulations built from living people’s intimate relationship archives

Article Title: How generative AI transforms relational archives into simulations of living persons

Article References: Luo, J. (2026). How generative AI transforms relational archives into simulations of living persons. Discover Artificial Intelligence, 6(1), Article 1161. https://doi.org/10.1007/s44163-026-02249-1

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02249-1

Keywords: generative AI, AI companionship, person-specific simulation, relational data, digital archives, ex-partner chatbots, synthetic disclosure, data consent, emotional attachment, AI governance, generative, transforms