With generative artificial intelligence and particularly beginning with the release of ChatGPT, many, many people have become very used to conversational AI experiences. A common part of prompting, is to assign a role or persona to the AI and instructing it to respond from that perspective.
This of course does not change the identity of the AI; it does change the words it uses in its responses. That can be helpful in terms of result quality and it can inform the nature of the experience. In that sense, the practice rewards experimentation.
Some people have asked their AI to assume the persona and communicate from the perspective of a deceased loved one.
As a concept this has received the attention of academic research and (of course) for-profit enterprises. Here is how Wikipedia introduces the idea of griefbots:
A deadbot, deathbot, or griefbot is a digital avatar, created with artificial intelligence, which resembles a person who is dead. Griefbots employ natural language processing and machine-learning techniques to approximate the style and personality of a deceased person. They may appear as chatbots, voice assistants, or animated avatars, and are often trained on an individual’s digital remains.
Wikipedia entry on Deadbot, retrieved on February 20, 2026.
With generative AI and the appropriate custom data, it has become technically possible to create conversational (and visual and auditory) experiences, such that it can have the (perhaps strong) semblance of interacting with that person.
This is – and in keeping with a recurring theme on this blog – an excellent example of the map and territory disconnect. If this is the experience you engage with, you are quite literally interacting with a(n AI) model, not the modeled. Though of course in this case, the map appears to be talking back, it is still just a map, albeit an impressive (and expressive) one.
Enabling people to be “communicating with dead people,” particularly deceased family members, was of course not the primary intended use case that led to development of generative AIs and chatting with large language models (LLMs), but here we are.
At the onset of a new technology it is rarely very clear what it will be used for. People are usually inventive at finding new and interesting applications that the original developers did not foresee.
People do and have come up with interesting applications for conversations with LLMs — and they survive to the extent that they are useful and provide value for people.
In a recent article for Scientific American, David Berreby discusses the phenomenon of griefbots, particularly in the context of people who are in grief, who are seeking to find comfort or find answers for unresolved questions. In light of the above, the following passage seems pertinent.
“It’s coming up in the lives of our clients,” [professor and therapist] Neimeyer says. “It’s an ineluctable part of the emerging technological and cultural landscape globally.” Whatever their views on the benefits and dangers for mourners, he says, “therapists who are consulted by the bereaved bear some responsibility for becoming knowledgeable about these technologies.”
Can AI Ghosts help us heal?
Technical possibility does not imply value judgment. Being able to do something is independent of whether you should be doing it.
This is possible and as Neimeyer points out above, people are doing this. It will take time and diligence to learn and understand whether, when and how it is a good idea.
Griefbots provide a model of deceased loved ones. Flawed and simplified, but interactive and likely useful for some people.
An important question is whether people, especially those in the throes of grief, are able to reliably distinguish between the model and the person it represents.
And how it matters, if they don’t.