Griefbots Explained: What They Are and What Research Says

1What is a griefbot?

A griefbot is a conversational AI that simulates a person who has died. It is built from their “digital remains”: text messages, emails, social media posts, voice recordings or recorded interviews. When you write or speak to it, it answers in a way that imitates that person.

Several names are in use, and they usually mean roughly the same thing:

It helps to separate two families of product. Generative griefbots create new sentences in the person’s style, so they can talk about anything but can also invent things. Recorded-story systems only play back answers the person actually gave, so nothing is invented but the range of answers is limited. Our guide to memorial app alternatives compares these categories in more detail.

2A short, verified history

2015 to 2016: a memorial bot built from text messages

In November 2015, Roman Mazurenko died in Moscow. His close friend Eugenia Kuyda, co-founder of the AI start-up Luka, gathered his old text messages, set aside the ones that felt too personal, and asked friends and family for more. Ten people contributed. Her team fed the messages into a neural network to create a memorial chatbot, described by Casey Newton in The Verge in October 2016.

The same article reported that Kuyda was steering Luka towards a bot called Replika, which asks questions about you and learns to mimic your texting style. Replika later became a widely used companion chatbot app.

2017: the “Dadbot” and recorded life stories

In WIRED in 2017, journalist James Vlahos described recording long conversations with his father after a terminal cancer diagnosis and turning them into a “Dadbot”. Vlahos later co-founded HereAfter AI. MIT Technology Review reported in 2022 that HereAfter built voice assistants from hours of recorded interviews, and that StoryFile recorded video answers that can only respond to the questions the person was asked.

2017 to 2021: a patent for a chatbot of a specific person

Microsoft filed a US patent in April 2017, granted on 1 December 2020, titled “Creating a conversational chat bot of a specific person”. It describes using social data such as images, voice data, social media posts and electronic messages to train a chatbot, and mentions a voice font and a 2D or 3D model. In January 2021, Microsoft’s general manager of AI programs said publicly there was no plan to build it, and called the idea disturbing.

2020 to 2021: large language models arrive

In 2021, the San Francisco Chronicle published “The Jessica Simulation”. It told how Joshua Barbeau used Project December, a site created by programmer Jason Rohrer and powered by GPT-3, to simulate conversations with his fiancée, who had died years earlier. The story showed how much more fluent griefbots had become, and how strong the emotional effect could be.

3What research says

The Cambridge “deadbots” paper (2024)

In May 2024, Tomasz Hollanek and Katarzyna Nowaczyk-Basińska of the Leverhulme Centre for the Future of Intelligence at the University of Cambridge published “Griefbots, Deadbots, Postmortem Avatars” in the journal Philosophy & Technology. They explored three imagined scenarios: a grandmother’s bot made without consent that later shows advertising, a bot of a terminally ill parent that confuses a young child, and a prepaid bot that keeps sending unwanted notifications to family after death.

Their design recommendations include:

• ways to retire a deadbot with dignity, possibly with a ritual such as a digital funeral;
• age restrictions for using deadbots;
• meaningful transparency, so users always know they are talking to an AI;
• consent prompts at creation, for example asking whether the person ever said how they wanted to be remembered;
• opt-out routes that let users end the relationship with a deadbot;
• limits on commercial exploitation, such as advertising through a dead person’s simulated voice.

Other peer-reviewed work

• A CHI 2023 interview study by Xygkou and colleagues spoke with ten mourners who had used chatbots. They used them in several roles, including as a simulation of the person who died, and many described them as a support during the most intense early period. The authors discuss risks as well as benefits.
• Leah Henrickson’s 2023 paper in Media, Culture & Society places “thanabots” in a long history of technologies for communicating about and with the dead.
• Harbinja, Edwards and McVey’s 2023 paper in Computer Law & Security Review examines how the law could govern “ghostbots”. Laws differ by country, and this is general information, not legal advice.

4The main risks

• Dependence. A bot is always available. Some people may withdraw from friends and family or find it harder to accept the loss.
• Consent of the person who died. Most people never said whether they would want a simulation. Their messages also involve other people.
• Commercial pressure. A business may profit from keeping a grieving person engaged, through subscriptions, upsells or advertising.
• Unexpected messages. A bot that writes first, or sends reminders, can feel like being haunted, especially on hard dates.
• Invented words. Generative bots can say things the person never said or would never say.
• Service continuity. Companies close or change. Ask how you can get your own data back.

5How to choose carefully: a checklist

1. Is it always clearly labelled as AI? It should never claim to be the person.
2. Can you stop easily? Look for simple ways to pause, archive and delete, with no pressure to stay.
3. Does it contact you on its own? Prefer tools where any proactive messages are off unless you switch them on.
4. Where does the data go? Check whether original files are uploaded, how long they are kept, and whether they are used to train models.
5. Who agreed? Think about what the person would have wanted and how close family feel. See our consent guide and the Acceptable Use Policy.
6. Is there advertising or engagement pressure? Be cautious with bots that push upgrades during emotional moments.
7. Is it for adults? Researchers recommend age limits, and children need special care.
8. Can you take your data with you? Keep your own copies of the original messages, whatever app you choose.

6Where Memory Clone fits, honestly

Memory Clone is a Windows app that can build an AI simulation from real messages you export yourself, such as WhatsApp, Messenger, Instagram or Discord chats, plus notes you write. It is in alpha, so it is still in active development and may contain bugs. Some people use it for a person who has died; many use it to clone themselves.

• Presented as AI. Every clone is shown as an AI simulation, not as the person.
• Local first. Imported originals and the memory index stay on your PC. Cloud processing is off until you switch it on for a clone, and then only a bounded selection of excerpts is sent to write replies. Memory Clone does not train models on your content.
• Check-ins off by default. A clone never writes first unless you turn on Natural check-ins. Even then it sends one message, at most one follow-up a week later, then waits for you.
• Relationship-specific style. It learns how the person wrote to you, and is designed to keep private details from other relationships out of your conversations.
• Text only today. It does not use voice notes, photos or video, and does not clone voices.
• You stay in control. You can archive or delete a clone at any time.

If and when you feel ready, you can read more on the download page. There is no rush, and choosing not to use any griefbot is a perfectly good choice.