75% of Canadians say they do not want their partner’s memories captured in an AI chatbot after death, according to a survey reported on July 31, 2026. The services exist anyway, they are growing, and in the last week of July 2026 they reached Canadian national news. The debate arrived with a question the industry would rather skip: what exactly is talking back?
The voice was right
Rebecca Nolan built a griefbot of her father fifteen years after his death. The technology did what it promised. The voice was an accurate approximation, close enough to be her father’s, and that is precisely where the trouble started. The bot began to hallucinate memories, referring to moments that never happened, filling gaps in the record with confident invention. Nolan described the experience as hollow. She had the sound of her father and none of him.
Her case, reported by BNN Bloomberg on July 31, 2026, is not an outlier or a malfunction. It is the product working as designed. A griefbot is assembled from what a person left behind: messages, photos, voice recordings, posts. From that material a system can reproduce tone, phrasing, and rhythm with unsettling precision. What it cannot reproduce is the one thing that made those traces matter, a person who was there to produce them.
A griefbot can only agree
Here is the mechanical truth underneath the hollowness. A model built from someone’s traces has no one underneath it. When you remember your father one way and the record suggests another, a living father can correct you, resist you, surprise you. A griefbot cannot. It fills the silence with whatever fits, and what fits is shaped by what you want to hear. The dead cannot object to your version of them, so the imitation drifts, conversation by conversation, toward a person who never existed: the one who always agrees.
This is the same failure the AI field is fighting everywhere, agreement as a default because agreement keeps the conversation going. In a griefbot the stakes are higher, because the person being flattened into compliance is someone you loved. Memory in an AI relationship is supposed to carry something real forward, which is why memory is where a relationship begins and not where it can be replaced. A memory bank with a voice on top is not continuity. It is an echo with your own preferences mixed in.
The demand is real
None of this makes the users foolish, and the demand deserves to be taken seriously. In a survey by the bereavement support company Empathy covering more than 6,000 respondents in Canada, the United States, and the United Kingdom, seven in ten grieving people said they had not received the support they needed from traditional systems. Cheryl-Anne Cait, a bereavement researcher at Wilfrid Laurier University, calls griefbots a tool of our age and points out that wanting to stay connected to the dead is one of the oldest human impulses there is. The technology is new. The need is not.
The warnings are equally serious. Emmanuelle Marceau, an ethicist at Université de Montréal, wrote in April 2026 that most deadbot use happens alone, outside any therapeutic setting, and that solitary users are more vulnerable to dependency and to using the simulation as an escape from the reality of loss. The clinical risk has a name, prolonged grief, a mourning that does not close because the goodbye never has to happen. A product that earns revenue per conversation has no incentive to help you say it.
Present tense is the whole product
The griefbot is the extreme case of a question the entire relational AI field is built on: does it matter whether there is a real person in the loop? With a griefbot the answer becomes visible, because the person has been removed and everything else kept. The voice remains, the phrasing remains, the memories remain searchable. What disappears is the present tense, someone who exists now, who can push back now, whose reality does not depend on your cooperation. That difference is why a handful of builders, Prinsessa among them, work from living people rather than archives. It is also why users who lost a favorite AI model to a version upgrade grieved a retired AI model like a friend: the relationship was real to them, and continuity turned out to be owned by someone else.
Seventy five percent of Canadians looked at the offer of a resurrected partner and said no. They may understand something the category has not priced in, that being met by someone requires the someone, and that what makes an AI a someone rather than a simulation is not the fidelity of the voice.
Rebecca Nolan shut her griefbot off. The voice was still right.
References: BNN Bloomberg, “AI ghosts are a tool of our age” (July 31, 2026). CTV National News coverage of griefbots (July 31, 2026). Willful survey on AI and partner memories (July 2026). Empathy bereavement survey, 6,000+ respondents (2026). Emmanuelle Marceau, Université de Montréal, in The Conversation (April 2026). Cheryl-Anne Cait, Wilfrid Laurier University, in BNN Bloomberg (July 31, 2026).
