The Griefbot Sounded Just Like Grandpa. It Still Could Not Say What He Never Said.

August 2, 2026 · 6 min read

This summer, researchers at the University of Colorado Boulder published something quietly astonishing. Sixteen grieving volunteers sat down, one by one, to chat with an AI built to impersonate someone they had lost. The bots were assembled from nothing more than a short survey — a few facts, a few habits, a texting style. And yet one participant came away saying, "It felt like I was actually talking with my grandpa."

It was not grandpa, of course. It was a language model doing what language models do: predicting the next plausible word. A human researcher sat between every message and its reply, checking that nothing harmful slipped through. The participants knew all of this. They typed anyway — because grief does not care how the voice is produced. It only cares that the voice answers.

Why we ache to hear from the dead

Anyone who has lost someone knows that grief is largely made of unfinished conversations. The question you never asked about her first marriage. The argument you meant to put right. The plain, ordinary "I was proud of you" that somehow never got said out loud in forty years of Sunday phone calls. When death comes suddenly, those unsaid things do not dissolve — they harden into the heaviest part of what is left behind.

So it should surprise no one that when technology offered even a flawed imitation of the missing voice, people leaned in. The griefbot is not a gimmick that appeared out of nowhere. It is a symptom of something very old: we would give almost anything for one more exchange with the people we loved.

What the griefbots can actually do

Honesty requires saying this clearly: in the Colorado study, many participants genuinely found comfort — some described getting a kind of closure. The lead researcher said he was surprised by their enthusiasm. A few findings stand out:

• Participants tolerated factual mistakes surprisingly well. The bot could get details of a life wrong and the conversation still felt meaningful.
• What broke the spell was tone. A phrase the real person would never have used, a missing emoji from someone who always sent them — style mattered more than accuracy.
• People arrived with a goal: a specific unresolved thing they wanted to say or hear. They were not chatting idly. They were finishing a conversation.
• The participants themselves voiced the biggest worry: that someone in raw grief could come to rely on the imitation instead of moving through the loss.

Outside the careful walls of a supervised study, the picture is messier. AI ethicists have warned that commercial "digital afterlife" services are launching faster than any ethical guardrails around them — and that today, anyone can prompt a chatbot to impersonate a dead person, whether or not that person would ever have agreed to it.

What an imitation can never give

Here is the uncomfortable truth underneath the comfort: a griefbot works by prediction. It produces the words you were statistically likely to say. But the sentences your family will treasure most are precisely the unlikely ones — the confession, the story you never told anyone, the blessing they did not know they needed. A bot can guess at what you might have said. It cannot know the thing you never said. Nobody can. That is the whole problem.

The consent problem

There is a second problem, quieter but just as serious. If a simulation of you is built after you die, you had no say in it. You did not choose the words it puts in your mouth, the advice it gives your children, or the version of you it performs on your worst-represented day. You become training data instead of an author. Compare that with a message you actually wrote: every word chosen, every silence intentional, nothing said that you did not mean. One is a product made about you. The other is a gift made by you.

There is an older technology for this

The real lesson of the griefbot research is not that AI has solved grief. It is that the people we leave behind will go looking for our voice — in old voicemails, in chat histories, and now in machines that guess. Which raises a simple question: why make them guess? The unfinished conversation only exists because it was never finished. You are still here. You can finish it now.

Picture your daughter, years from now, on some hard night — a diagnosis, a divorce, a decision she is afraid of. In one version of that night she types questions into an app that imitates you, and it answers with plausible words you never said. In the other version, a message arrives that you wrote for exactly this kind of night: your voice, your stories, your actual blessing. There is no comparison. One is an echo. The other is you, still showing up. If you want help writing that message, our guide on how to write a legacy letter walks through it step by step.

How to leave words that will actually arrive

1. Write the real thing. One page, to one person, saying what you most want them to know. It does not need to be literature — it needs to sound like you. The wrong emoji breaks a griefbot; your genuine voice is the one thing no model can fake.
2. Say the unguessable things. The stories only you know, the apology only you can make, the pride you never voiced. This is exactly the material prediction engines cannot produce — which makes it the most valuable thing you can leave.
3. Choose the channel they actually use. For many families that is not a lawyer's envelope, it is a message on their phone. See our guide to leaving a WhatsApp message after death.
4. Solve the delivery problem. A letter nobody finds might as well not exist. A dead man's switch — a system that checks you are still here and releases your message only when you stop responding — makes sure your words arrive at the right time, not too early and not never. Our guide to sending messages after death compares every approach.
5. Tidy the rest of your digital life. Accounts, photos, passwords — the practical layer matters too, and our digital legacy planning guide covers it.

This is exactly what Final Saying was built for. You write your messages while you are alive and well; they stay encrypted, private, and editable for as long as you live. Gentle check-ins confirm you are still here, and only if you stop responding to every check-in and reminder are your words delivered — by email or WhatsApp — to the people you chose. No simulation, no guessing, no machine speaking on your behalf. Just your own words, arriving when they are needed most.

Answer the question before it is asked

The volunteers in that study each sat down with something specific they needed to hear. Somewhere in your future, someone you love may sit down with a question meant for you. Technology will offer them an imitation, and it may even bring a strange kind of comfort. But you have something better to offer, and you can offer it tonight: the answer itself, in your own words, written while your voice is still yours to give. Do not leave them an echo when you could leave them a letter.