Should you use AI to talk to someone who died?
We built Still Here so it will never impersonate a person who died, because grief research distinguishes healthy continuing bonds, carried in memory and identity, from relating to the deceased as still externally present and responsive, a pattern linked with poorer adjustment. A griefbot manufactures exactly that external presence, on demand, indefinitely.
The pitch is understandable, and that is what makes it worth taking seriously. If you could hear one more "goodnight" in his phrasing, ask her the question you never asked, would you? Many people would, and an industry now exists to sell it: systems trained on a person's messages that generate new conversation in their voice.
What the grief research actually distinguishes
Since Klass, Silverman and Nickman's 1996 work on continuing bonds, it has been mainstream in bereavement science that staying connected to the dead is normal and often healthy. The relationship doesn't end; it transforms. People carry a parent's values, talk aloud to a spouse at the graveside, feel a presence in a loved room. None of that is pathology.
But the research draws a line inside that finding. Field and colleagues, reviewing the evidence in 2005, distinguished internalized bonds (the deceased as inner presence: memory, identity, legacy) from externalized expressions that locate the person as still concretely out there and interactive. The internalized form tends to travel with better adjustment; the externalized form, particularly when it substitutes for taking in the reality of the death, tends to travel with worse.
Why a simulation is not a photograph
Photos, voicemails, and letters are relics: finite, fixed, and honest about being the past. A griefbot is generative. It produces new responses, which means it does not commemorate the relationship so much as continue it, with a counterpart who never tires, never changes, and never says the difficult thing. Grief involves, among much else, the slow integration of an unbearable fact. A system whose entire function is to behave as if the fact were not true is pointed in the other direction, and it is available at 2am, every night, for a monthly fee. The commercial incentive and the clinical risk point the same way.
The boundary we chose, and what we do instead
Still Here refuses, at the architecture level, to speak as the person you lost: no messages from them, no roleplay, no "what would they say." A separate output filter enforces it even if the conversational model slips. What it will gladly do is help you remember them in your own voice: tell their stories, write TO them rather than receive letters from a simulation, carry forward what they gave you, and sit with what was left unsaid. That is the form of connection the evidence supports, and it keeps the one thing a simulation quietly takes: the truth.
References: Klass, Silverman & Nickman (1996) Continuing Bonds. Field, Gao & Paderna (2005) Death Studies 29(4). Stroebe & Schut (1999) Death Studies 23(3). See also emerging HCI and ethics literature on "deadbots" and digital remains.