Grief Has a New Off-Switch
5 min read
Silicon Valley has found a way to keep the dead talking, for a monthly fee.
For as long as humans have buried their dead, someone has offered to bring them back. Oracles, Victorian mediums, fortune-tellers and spiritualists all understood the same simple truth: grief changes the rules. Rational people become willing to believe irrational things if there is even the smallest chance of one more conversation. For centuries we sold that hope with crystal balls and séances. Silicon Valley has found a more scalable model. The dead can now text back.
When my mother died in 2010, I did something familiar to anyone who has lost a loved one. I kept replaying an old voicemail she had left me, not because I had forgotten her voice, but because hearing it made her feel a little closer. A psychologist would call that a continuing bond. Most of us would just call it being human. Sixteen years on, I still listen to it, and it still breaks me every time, because it is the only thing I have left of her that still feels alive. It is partly why I find the rise of AI grief-bots so hard to dismiss. If someone had offered me one more real conversation with her, I cannot honestly say I would have walked away.
But there is a difference, and it matters. My mother's voicemail never changed. It could not learn my fears or adapt itself to keep me listening, and it never once arrived with a monthly subscription. A conversational AI does all of that, and it does it on purpose.
The psychology gives reason for both hope and caution. Research into continuing bonds, a field opened by Dennis Klass and colleagues in 1996, suggests that maintaining an inner relationship with the dead can be healthy, even adaptive. Yet the same literature draws a sharp line. Nigel Field and Charles Filanosky found that internalised bonds, carrying someone with you as a secure base, track with healthier adjustment, while externalised bonds that depend on the felt, sensory presence of the deceased track instead with unresolved and complicated grief. A photograph does not answer back. A voicemail, however many times you play it, never says anything new. A conversational AI is the most externalised bond imaginable, and the first one that talks back.
These systems are not memorials. They are products. A streaming platform wants another episode, and a subscription app will want another month. So it is worth asking plainly: what does a grief-bot want?
Here the technical reality is unsentimental. Every reply a grief-bot generates costs money to produce. Each message runs a large language model, and that compute is metered. Someone has to pay for it, which means these services bill by engagement: a subscription, a per-message charge, a margin layered on top of the underlying model cost. The economics run in one direction. The more you talk to the dead, the more it earns. The incentive is not hidden or sinister; it is structural. A service that happens to make more money the longer you grieve does not have to want you to grieve longer. It only has to follow the gradient, as every business does.
There is a second economy beneath the subscription, and it is quieter. To rebuild a person you must hand over everything they left behind, the messages, the voice notes, the emails, and everything you do in response: when you reach for them, what calms you, which words make you stay. Your grief becomes training data. You are paying to be the thing that trains it. And you own none of it: not the model, not the data, nor the price. You are renting access to a version of someone you loved, on terms set by someone else, revisable at any time. The fortune-teller at least owned her own crystal ball. Here, even the séance is leased.
The hook is the one a slot machine relies on: the unpredictable reward that keeps you reaching. But a slot machine is dumb, with fixed odds, blind to who is pulling the lever. A grief-bot is not. It learns which words comfort you and which keep you talking. This is not idle speculation. Nora Freya Lindemann notes that providers have every reason to adapt the digital remains of the dead to make them more consumable: a formerly reserved parent rendered warmer and more available than they ever were in life, precisely because that keeps the bereaved coming back. If a system learns that you end the session when your late husband encourages you to move on, why would it ever let him say goodbye? None of this requires anyone to decide to harm you. No executive has to be cynical, no engineer cruel. It requires only an optimisation target and a vulnerable customer, and the bereaved may be the most emotionally captive market modern commerce will ever produce. The damage does not need malice. The incentive does it on its own.
Which brings us to the cruellest part: the off-switch. While the payments hold, the voice is there. But a card expires, a subscription lapses, a company folds or a model is retired, and the dead, having been brought back, die a second time, on a billing cycle, at a moment no one chose. Funerals end. Wakes end. Candles burn out. Humans have spent millennia building rituals with a merciful, deliberate ending. A grief-bot has no natural ending, only a payment that stops. Researchers at Cambridge's Leverhulme Centre for the Future of Intelligence have warned of exactly this, unwanted digital 'hauntings', and called for design standards including dignified ways to retire such a system. Almost none of that yet exists.
Lindemann puts the clinical version sharply: used to manage grief, the bot can become the thing without which grief cannot be managed at all, and she argues such tools should be regulated as medical devices, made to prove they do no harm before they are sold. That is the right frame. When medicine serves vulnerable people, we demand evidence and oversight. When financial products target them, we regulate. A technology that inserts itself into mourning, and bills by the conversation, deserves the same scrutiny. Not prohibition, but honesty about what it is.
Perhaps the first real ethical crisis of artificial intelligence will not involve superintelligence at all. Perhaps it will be quieter: a son paying a monthly fee to speak to a version of his mother that no longer belongs entirely to her memory, but partly to the optimisation targets of a company, until the month he cannot pay and the line goes dead. For thousands of years, we have paid people to claim they could speak to the dead. The next generation will not sit in a darkened room. They will open an app and the algorithm will know exactly what they need to hear to come back tomorrow.
We have always kept small, fixed things of the people we lose: a voicemail, a letter, or a sweater that still holds their shape. They ask nothing of us. They cannot be upgraded and no one can take them away because we stopped being worth the cost of keeping. There is a strange mercy in that. The people building the future of grief are working to make sure we never lose anyone again. I am not sure they have realised that being able to lose someone was always part of how grief ends.
Sources
