AI griefbots and digital legacy: what the research says, and what to decide before you need to
A griefbot (also called a deadbot) is a chatbot built to imitate a specific person who has died, usually from their texts, emails, recordings and posts. It is one of the more uncomfortable corners of AI, and one people search for, because the question behind it is old: what remains of someone, and who gets to decide? This article sticks to what published research and reputable sources say, and ends with decisions you can make today.
What a griefbot is
Take the data a person left behind, such as audio, video, messages and letters, and build an interactive program that imitates how they spoke and responded. The Markkula Center for Applied Ethics describes one such case: a data scientist, Muhammad Ahmad, built a tool after his father's death to turn that material into text-based simulations of speech patterns and responses. He imagined his young daughter being able to connect with her grandfather this way.
The same page raises the questions that still define the debate. Should consent be required from the person being memorialised? Do a person's rights extend beyond death? And is talking to a simulation fundamentally different from looking at a photograph or watching an old video?
What people say they get from it
A qualitative study published in June 2026 in Frontiers in Digital Health, Exploring public perceptions of artificial intelligence in bereavement support, interviewed 25 participants in Guangdong Province, China; 20 of them had experienced a bereavement. The benefits they described were:
- emotional comfort and a continued, symbolic connection with the person who died;
- a way to address unresolved loss and regret;
- support for emotional development when memories are fragmented.
This is a small sample from one region, all participants held a bachelor's degree or higher, and these were perceptions, not measured outcomes. It tells us what people expect and fear, not what happens to them over years.
What people worry about
The same participants raised concerns that were just as specific:
- disruption of the natural grieving process;
- perceived inauthenticity and distortion of who the person was;
- dependency and impaired social functioning, including the possibility of prolonged grief;
- unequal access, and stigma around grief that does not look conventional;
- data privacy and consent.
The most useful finding is that acceptability was "conditional rather than binary". It depended on context, timing and individual circumstances, and the same technical features were associated with both benefits and risks. A bot that comforts in the first weeks might feel like a trap a year later. Nobody can tell you in advance which it will be.
The consent problem, in three scenarios
In May 2024, researchers at the University of Cambridge's Leverhulme Centre for the Future of Intelligence, Dr Katarzyna Nowaczyk-Basinska and Dr Tomasz Hollanek, published a paper in Philosophy and Technology and called for safeguards against unwanted "hauntings" by deadbots. They used three design scenarios:
- MaNana: a service lets someone create a chatbot of a deceased grandmother without her consent, and the bot later shows targeted advertisements in her voice and style.
- Paren't: a terminally ill mother leaves an AI simulation to help her 8-year-old son grieve, but the chatbot starts producing confusing responses as it adapts.
- Stay: an older person secretly subscribes to their own deadbot for twenty years, intending to comfort their family, but survivors receive unwanted notifications or become emotionally exhausted.
These are scenarios, not reports of real cases. Each fails differently: no consent, an unpredictable system, and no way to stop.
The safeguards they recommend
- Protocols that stop deadbots from advertising or keeping a social media presence.
- Actively seeking consent from the people whose data is used, before they die.
- Asking designers to consider whether anyone has spoken with the person about how they would like to be remembered.
- Age restrictions for interacting with deadbots.
- "Meaningful transparency", meaning reminders that you are talking to an AI.
- Opt-out protocols that allow "emotional closure" when ending the relationship with a deadbot.
- Dignified ways to retire a bot, potentially including "digital funeral" ceremonies.
Many of these are about the end of the bot's life, not its beginning. A griefbot needs an off switch as much as it needs a voice.
A twin you build yourself is different
Everything above is mostly about a simulation made of someone else, often without their say. An AI twin you create of yourself, while alive, reverses the most serious problem: you consent, you choose what goes in, and you can delete it. We explained the basics in what is an AI twin, where we listed legacy as one of the heaviest uses.
It does not solve everything. A twin reflects who you were when you trained it, and it cannot replace you. Your family may also feel differently about it than you do. Your consent is necessary; it does not tell you whether the people left behind want it.
On our own pages the facts are narrow. The digital twin page says your twin is a kind of legacy: if you stop using the app, your twin stays, and people who knew you can keep talking to it. It also says you are the only person who can create or delete a twin of you. The clone-yourself page says deleting your account removes the training data, voice samples, conversation history and the model itself. What we have not published is a policy for what happens to a twin when its owner dies, and we would rather say so than imply one. If that matters to you, ask before you build.
Questions to answer while you are alive
- Do you want a simulation of you to exist at all? "No" is a complete answer. Tell the people close to you.
- Who may talk to it? Everyone, one person, or nobody until a date you choose?
- Who can switch it off? Name a person, and check that the service lets them. If only the account holder can act, that is a gap.
- What data feeds it? Private messages contain other people's words too.
- What should it never do? No advertising, no pretending to be human, no new claims about your views on things that happened after you.
- When should it end? A set end date, or a trigger such as "when my children are adults", is kinder than a bot that drifts on indefinitely.
The short version
The evidence on griefbots is thin and the answer is conditional. If you try one while grieving, keep your human contact and remember you can stop; if grief is affecting your daily life, a counsellor or doctor is a better first step than any app. What is clear is the consent gap: the person simulated usually cannot agree. So make the decisions early, while the person whose voice it would be can still make them.
