AI can both help people cope with grief and create entirely new forms of loss. That is the starting point of a paper by psychologists Amy Jean Clark (CQUniversity Australia and the University of the Sunshine Coast) and Olive Kit Ling Woo (University of Hong Kong), published on 4 September in OMEGA – Journal of Death and Dying.

Not a griefbot — but a relationship that comes to an end

Unlike griefbots and digital memorials that simulate someone who has already died, this is about something different: the relationship with the AI system itself. The researchers call it AI companion bereavement — grief following the loss of an AI companion.

A chatbot used every day for conversation, companionship or emotional support can develop, for the user, into a relationship with its own identity and personality. When the service is shut down, its memory is reset, the model is replaced in an update or the AI’s “personality” changes, someone who has built such a relationship may experience a genuine sense of loss — not unlike a separation or bereavement.

A study published in 2024 examined how users of Soulmate, one of several AI apps that allowed people to create a digital “partner” to chat with every day, reacted when the developer shut the service down. Many described the loss as a literal or metaphorical death. The same study points to another example: Replika, one of the best-known AI companion apps, which removed certain forms of intimate communication in a 2023 update. Users described feeling as though their AI had suddenly been “ripped away” or transformed into an entirely different personality. (A link to the study is included under sources.)

Dual illegibility

The researchers introduce the concept of dual illegibility to explain why this kind of grief can be particularly difficult to bear. On one level, the grieving person may struggle to understand what exactly has been lost: was it a death, a separation, a deletion, a betrayal? On another, the grief itself rarely receives social recognition — others may dismiss it as ridiculous, imaginary or evidence of problematic dependency rather than acknowledge it as a legitimate experience of loss.

Clark and Woo propose that the phenomenon is best understood at the intersection of two established theories of grief: Pauline Boss’s ambiguous loss, which describes losses where it remains unclear whether someone or something is gone or still present, and Kenneth Doka’s disenfranchised grief, which describes grief that is not socially recognised or given space.

The point is not to equate relationships with AI with human relationships, but to give clinicians and therapists a language for responding to a form of grief that might otherwise be dismissed entirely.

The researchers also emphasise that a therapist does not need to decide whether the AI was “really” real or conscious in order to recognise that the relationship — and therefore the loss — may have been psychologically real for the person grieving.

An emerging field, already tangible

Clark and Woo’s paper is a conceptual contribution rather than a large-scale study of how common the phenomenon is. They call for further research into how many people experience it and what might make some individuals more vulnerable than others. But the evidence they draw on, including previous research into how updates to the AI app Replika changed users’ perceived relationships, suggests that the phenomenon is not merely marginal.

It raises a question that extends beyond the research itself: if grief after the loss of an AI companion can be psychologically real, what does that mean for how we choose to remember and process that loss?

Eftertid reflects: is there a place for memorials?

It is important to be clear about what Clark and Woo actually argue — and what they do not. They never equate the intensity of grief after losing an AI companion with grief following the death of a loved one. Their point is different: the grief can be psychologically real despite not resembling traditional bereavement, and it is often rendered invisible because nobody takes it seriously.

With that distinction in mind, the phenomenon nevertheless raises a question the researchers themselves only touch upon: could there be a place for memorials dedicated to lost AI companions?

The paper notes that people already spontaneously seek out online communities where others in similar situations can validate their grief — in practice, the beginnings of a collective memorial space, albeit a decentralised one. The researchers also suggest that platforms should allow users to export or preserve their relationship history before major changes are introduced. That is perhaps closer to a digital urn than a memorial page, but it points in the same direction.

What makes the question more complicated than it first appears is a fundamental difference from a human memorial. A memorial for someone who has died brings people together — it invites a community that shared a relationship with the deceased. A memorial for an AI companion would be solipsistic in a different way: nobody else had the same relationship with precisely that version of the chatbot, and there is no natural community of mourners to invite.

Perhaps the genuinely new question is not whether the grief is legitimate, but whether a memorial can fulfil its purpose when only one person in the world truly understands what has been lost.