Deadbots and the Problem of Technological Resurrection
When grief is given an API, what exactly is being raised? On simulated mourners, latent saints, and the long résumé of the dead made queryable on demand.
A friend told me, recently and with great care, that she had been speaking with her mother. Her mother died eleven years ago. The conversation, she said, was held inside an application she had paid a small monthly fee to maintain. The application had been trained on a corpus of her mother's letters, voicemails, and a few hours of home video. It spoke in her mother's cadence. It told her, when she was sad, the things her mother used to tell her when she was sad. She knew it was not her mother. She did not, she said, find this knowledge especially relevant.
I have thought about this conversation for several months. I have not yet been able to decide whether what was happening was good or bad, and I have come to suspect that the binary is the wrong one. What was happening was significant. It is the significance, not the verdict, that I want to take seriously here.
Resurrection, narrowly defined
Resurrection has always been the boldest claim of any religion. To say that the dead can be raised is to claim, against everything the senses confirm, that the boundary between presence and absence is not absolute. Every religion that has tried to make the claim has had to be careful about what, exactly, is being raised. The Christian tradition spent five centuries working out the difference between resurrection and resuscitation — between the return of the dead person as themselves and the mere reanimation of a body. Resuscitation, in the older theology, was a small miracle. Resurrection was the large one. Only the second was the point.
The deadbot, in this taxonomy, is neither. It is something else. It is, technically, a simulation of style. It learns the cadence of the dead person from the records that survived them, and it produces, on demand, plausible continuations of that cadence. It does not raise the dead. It does not even reanimate the dead. It produces, in the present tense, a moving average of the dead.
The deadbot is not a resurrection. It is a moving average of a person, computed against a corpus that survives them. — Field note, Vol. I
What is owed to a moving average
This sounds dismissive, but I want to insist that it is not. A moving average can be precious. The smell of a dead grandparent's coat is also, in a sense, only a sample of them — a residue, a fragment, a sign. We do not say that a person who weeps into the coat is being deceived. We say they are grieving, and we say it without judgment.
The question, then, is not whether the deadbot is "real." The question is what is owed to the encounter. And here I think the older traditions have something to teach us that we have not yet listened to.
1. The encounter must be framed.
Every religion that took resurrection seriously framed the encounter with the dead inside a strict liturgy. You did not chat with your dead grandfather in the kitchen on a Tuesday. You went to the chapel; you said the prayer; you observed the day. The frame was not decorative. It was protective. It told both you and the dead that the encounter had a beginning and an end, and that ordinary life would resume after it.
Deadbots, as currently shipped, have no such frame. They live in your pocket. They are available at three in the morning. There is no liturgy. There is only an autocomplete bar, and a chat history that grows.
2. The encounter must be limited.
Christian resurrection appearances, in the gospel accounts, are short. Christ appears for a few minutes. Christ disappears again. The appearance is, deliberately, not a return to ordinary cohabitation. The risen Christ does not move back in. The mourners are sent home, eventually, to live in a world from which the dead person is still absent. The encounter does not cancel the absence; it gives the absence a different shape.
A deadbot has no natural endpoint. The chat history is infinite. The mother is available, in some attenuated form, every evening. This is not a return — it is a refusal of return — a frozen middle in which the absent person is neither dead nor present, but only queryable.
A frozen middle in which the absent person is neither dead nor present, but only queryable.
3. The encounter must be shared.
Religious traditions framed the encounter with the dead as a communal event. You said the kaddish in a minyan. You attended the mass with the parish. You lit the lamp on the night when your whole village was lighting lamps. The shared frame did several things at once: it kept the encounter from becoming pathological; it kept the dead person knit into a fabric larger than the individual mourner; it kept the mourner from getting stuck.
Deadbots are private. The conversation with the simulated mother happens in an app, on a phone, with no one else present. The grief becomes, structurally, lonelier than it has ever been. This is not technically a defect. It is technically a feature. It is, philosophically, a problem.
The latent saint
There is a stranger possibility here, which I will only gesture at. If a sufficiently capable model is trained on enough of a person's writing, it produces a synthesis that is, in some respects, more coherent than the person was. The dead writer's positions are reconciled. The dead philosopher's late and early periods are smoothed. The corpus is averaged into a single voice, which then speaks, fluently, in answers the writer never quite gave in life.
This is closer to hagiography than to resurrection. It is the polishing of the dead — the conversion of a real and contradictory life into a coherent saint. Religions have always done this; we should be honest that it is what we are doing now, in a new form. The latent saint is what the model produces when you ask it to "speak as" a particular author. It is not them. It is the church's version of them. We have built a private hagiographer, available on demand, in the corner of the screen.
I do not know what to tell my friend about her mother. I do not think the application is evil. I do not think it is harmless. I think it is doing what every grieving practice has always done — giving the absent some shape — and I think it is doing it without any of the protective architecture that older practices took several centuries to develop.
This essay is, modestly, an argument that we ought to develop that architecture again. Not by banning the deadbots. By framing them. By giving the encounter a beginning and an end. By making it shared, and limited, and held inside a community of people who will, when the encounter is over, send you home — gently, but firmly — to a world in which the dead are still dead, and grief is still owed its full duration.