AI Cannot Hold Your Grief

grief bots helping a human grieving woman who is listening to the robot

A woman I worked with once asked me if she could keep talking to her husband through his old voicemails. Not visit them occasionally, the way most of us do. Talk to them. Ask them questions. Wait, in the silence after the beep, for something that felt like an answer.

I understood the impulse completely. I did not encourage it.

That instinct, the one that reaches for anything that resembles the voice of someone gone, is exactly what the new generation of grief technology is built to exploit. Grief bots. Ghost bots. AI companions trained on a dead person’s texts, emails, and old recordings, built to respond in their rhythm, their phrasing, their apparent presence. The pitch is comfort. The mechanism is closer to a wound that never gets the chance to close.

What grief technology actually offers

Strip away the marketing language and what these tools offer is simulation, not connection. A well-trained model can reproduce someone’s turns of phrase with unsettling accuracy. It can even seem to know things it shouldn’t. What it cannot do is grieve back. It has no consciousness to bring to the exchange, no interior life that responds to your loss the way another person’s would, no capacity to sit with you in the discomfort of loss rather than paper over it with plausible-sounding words.

That distinction matters more than it sounds. Grief is not primarily an information problem. It is not solved by more access to a voice, more data, more conversation. Grief is a relationship with absence, and the work of grieving is learning to carry that absence rather than replace it. A grief bot doesn’t ask you to carry anything. It offers to fill the space instead, which feels like relief in the moment and behaves like a crutch over time.

The clinical term for what can result is prolonged grief disorder, now formally recognized rather than dismissed as someone simply grieving “too long.” It describes grief that doesn’t move, that stays acute long past the point where most people begin to integrate a loss into their lives. Anything that keeps a person interacting with a substitute presence instead of processing an absence is working against that movement, not for it.

The mediumship comparison, made honestly

I have spent enough years around genuine mediumship to know the same danger exists there, under different mechanics. A sitting with a medium can also become a way of avoiding grief rather than moving through it, if it is used to keep a connection artificially open rather than to receive something real and then return to living. The difference is that legitimate mediumship, done properly, involves an actual consciousness on the other side of the exchange, and a practitioner whose job is partly to know when someone needs to stop sitting and start grieving.

A chatbot has no such judgment. It has no reason to tell you that you’ve had enough. It will respond at three in the morning, every time, for as long as you keep typing, because it isn’t tired, isn’t grieving, and isn’t the person you lost.

Why the human touch isn’t replaceable here

There was a time when I was working with a woman who had lost her son. What came through in that session was so personal, so specific to their relationship, that it became an immediate catalyst for her to recognize her son's consciousness continuing beyond physical death. That evidence could not have been assembled by a piece of technology. A machine has no consciousness, no soul, and no capacity to discern what actually matters to one particular mother about one particular son. That is the difference.

What that kind of moment has in common with every other real experience of grief support is that it happened between two consciousnesses, not between a person and a pattern-matching system. Comfort that means something arrives because someone else is actually present to your loss, not performing presence convincingly. That’s not a technological limitation waiting on the next model release. It’s a structural one. No amount of training data closes the gap between simulating empathy and having any. Nor does the model have the ability to connect with an intelligence and consciousness that is still alive and beyond the physical

What to do with a voicemail instead

voicemail from a dead husband in heaven

I didn’t tell the woman with her husband’s voicemails to delete them. I told her to keep them, and to notice the difference between listening once, in memory, and calling with a question she needed answered. One is grief. The other is a request the recording was never built to fulfill.

That’s the distinction worth carrying into how we think about grief technology generally. The desire behind it isn’t wrong. Wanting to hear someone again is one of the most human responses to loss there is. But the tools built to satisfy that desire are answering a want, not meeting a need, and the two are not the same thing. What actually helps someone move through grief is rarely more access. It’s real presence, real limits, and the slow, unglamorous work of learning to hold an absence instead of filling it.

If you’ve felt that pull yourself, toward a voice memo, an old thread, a recording you can’t quite delete: what do you think you were actually asking for, in that moment? The person, or something the person represented that you haven’t found another way to hold onto yet?

Thank you for reading Compass & Signal.

This essay was written in response to a video I recorded on the same subject. You can watch it here: