What Is an AI Avatar of a Deceased Loved One, and What Is It Actually Made Of?

An AI avatar of a deceased loved one is a computer persona that answers in their style, from the memories, recordings and conversations the family gathered. It is not the person, not a consciousness and not a "ghost". It is a language model that is only allowed to speak from what it was given. This article explains what the persona is actually made of, what it knows and does not, how much time and material it takes, and what happened in this field over the last two years. Without promises the product cannot keep.

What is it made of, in practice?

Three parts, and it is worth knowing all three by name:

• A general language model. The conversation itself is run by a model from Anthropic (Claude), the same kind of model behind the well-known AI assistants. It was not "trained on Grandpa". It knows how to phrase, understand a question and answer, and without the next part it would answer like any AI assistant.
• "Personality DNA". This is the part built from your materials. A separate model reads everything the family gathered, photos with the stories behind them, voice messages, a WhatsApp export, memories family members wrote, and extracts four layers: who the person fundamentally was, how they acted and reacted, how they spoke (recurring phrases, humor, rhythm), and which memories define them. That profile is handed to the conversation model before every conversation. The more real material there is, the more accurate it is; without material there is no persona, and a responsible system does not invent a personality from imagination.
• A reconstructed voice (optional). If recordings were kept, about a minute of clear speech is enough to start, even as a collection of WhatsApp voice notes, an external voice provider (ElevenLabs) learns the voiceprint and reads the answers in it. We went deeper in Voice Cloning of a Loved One.

And above all three, a fixed layer of rules that loads before the DNA, not after it: the persona always identifies as AI, never invents details that were not provided, and never discusses the circumstances of the death. Those rules cannot be switched off, not even by the family.

What does a conversation look like?

An example, invented for the sake of explanation, but this is exactly the shape:

Question: "Grandpa, what would you say about Danny moving abroad?"
Answer: "Well, what can I tell you. I left home at nineteen myself and never looked back. Let him go. Just let him call on Friday, that is all I ever asked of you too."

Notice what happened here: the persona does not "know" who Danny is or what is happening today. It leans on two things that were given to it, the story about leaving at nineteen, and the habit of asking for a Friday call, and phrases them in the rhythm learned from WhatsApp. If the story about leaving had not been in the materials, the answer would have been more generic. And that is fine: a generic answer is better than an invented memory.

What it knows, and what it does not

• ✓ Answers in the style, expressions and humor learned from the materials
• ✓ Tells the stories and anecdotes the family provided, and connects them
• ✓ Responds according to the values and patterns that were documented ("what would he say about…")
• ✗ Knows nothing about what was not given to it, no events, no people, no "secrets"
• ✗ Does not remember previous conversations, every conversation is independent, on purpose: so as not to cultivate an accumulating "relationship" that encourages dependence
• ✗ Does not talk about its death, does not impersonate a living person, and does not replace human connection or grief support

So how is this different from a regular chatbot?

Not in the model, in the limits. A regular chatbot answers from everything it ever learned. The avatar is guided only by the memories of one specific person, and is not allowed to go beyond them. Honesty requires saying that this is a constraint on a general model, not a new creature, and that the quality of the persona is exactly the quality of the material the family gave it.

On names: the press sometimes calls this a griefbot or a deathbot. Both describe the same thing, a chatbot built from the materials of a person who died, but they were coined mostly by critics, and it shows. We prefer "persona" or "avatar", not for marketing reasons, but because "bot" suggests something that answers everything, and a good persona is the exact opposite: it knows only what the family told it.

How long, and how much material?

A first conversation and a few photos, half an hour. The first DNA build, a few hours. A persona that truly feels like the person, a few weeks, at your pace, because every new piece of material sharpens it. The strongest materials, in order: a WhatsApp export (how to save it), voice messages, and photos with the story behind them (how to gather them from the whole family). The full process, step by step, is described in How It Works.

What happened in the field in 2024–2026

This category is young, and turbulent. In 2024 a University of Cambridge team called for regulation, including transparency requirements and protection for vulnerable users (the full announcement). In November 2025 an American company launched an app that generates a video persona from a few minutes of footage, and within days drew the sharpest ethical backlash the field had seen. In July 2026 HereAfter AI, the best-known product in the category, shut down, and its users were left asking what happens to their recordings (we wrote about it).

Those three events are why the rules above are fixed and cannot be switched off, and why everything can be exported and deleted at any moment. Anyone choosing a platform in this field should ask the export question first, not the feature question.

Is it healthy?

For some people, yes, in moderation. For some, it can do harm, especially in fresh grief. The research is young, and we gathered what it does know, with links, in Digital Legacy and Grief. Our bottom line: a tool for memory, yes; a substitute for parting or for professional support, no.

FAQ

Rather see than read? A live memorial page a living album and a persona you can talk with. What that conversation really is, and what it is not, in Talking to a deceased loved one through AI. And on the boundaries we set, and why, on our ethics page.