Are We Creating Digital Memorials… or Digital Mediums?
When an AI chatbot imitates the voice of someone who has died, what happens when grief makes it feel uncannily real?
Imagine being able to speak to somebody you love who has died.
Not just through a photo, an old voicemail or a letter tucked away, but through something that replies. You send a message, and it answers in a way that feels familiar. Maybe it uses a nickname you know or asks a question that person would have asked. For a moment, it almost feels like hearing them again.
It is not hard to see why someone might want that.
Grief often leaves us with unfinished conversations. We think about things we wish we had said. We wonder what the person would think of our new job, our kids, moving house, a decision we’ve made, or the challenges we’re facing now. Sometimes we know what they would say. Sometimes we just want to hear it anyway.
A recent study titled ‘Designing Conversations with the Dead: How People Engage with Generative Ghosts’ looked at what happened when 16 people talked to AI-modelled versions of loved ones who had died. The researchers developed two kinds of chatbots, including one that talked about the deceased in third person, and another that spoke as if it were the person themselves.
Some participants found the experience comforting; some felt close to the person again. One described receiving a form of closure.
These reactions are striking and are likely to be the part of the study that draws the most attention.
But the research doesn’t show that AI helps people grieve in a general or clinical way. Sixteen volunteers is a small sample, and everyone had already agreed to try talking to an AI version of someone they lost. People who found the idea upsetting or unacceptable weren’t included.
The research used brief, closely supervised conversations between participants and the chatbot. There was a facilitator, too: a researcher who reviewed and sometimes edited the chatbot’s replies before sharing them with the participant. This wasn’t someone alone at home, using an unfiltered bot during the hardest part of their grief, which is how such products are likely to be used in the wild.
Still, the study gives us something to think about. It shows how quickly a program built from just a few memories can start to seem like more than a memorial.
It can start to feel like you’re really talking to the dead.
Why would anyone want this?
People often talk about griefbots as if those who use them don’t really understand what they’re doing.
That doesn’t seem fair.
Many of us keep relationships with people who have died. We talk to them in our minds, visit graves and share our news. We keep their things, cook their recipes (or try to…), and imagine the advice they’d give. Someone can know their mother is gone and still think, “Mum would have loved this.”
Grief doesn’t end the relationship; it just changes how it continues.
In my last article about digital ghosts, I compared griefbots to rituals like wind phones and letters sent to the dead. These practices give people a way to express thoughts that still need to be said. They let the bereaved speak without pretending to get a reply.
AI shifts that arrangement.
The machine answers back.
That doesn’t make someone foolish for finding comfort in the answer. When we’re grieving, comfort doesn’t always depend on believing something is literally true. Someone might feel better after writing a letter they know won’t be read. They might dream about someone who has died and wake up feeling like they spent time together.
The emotional experience is real, even if we know the encounter is imagined, symbolic, or artificial.
What is different is that the machine begins doing some of the imagining for us. The grieving person is no longer only remembering what their loved one said, but the software is producing new sentences they never spoke, or maybe never would have.
In my previous article, I outlined why I find that incredibly chilling and, after writing that piece, I deleted my deceased mother’s old social media accounts. We had memorialised her Facebook account, and I realised that I don’t know if she’d have wanted that. The idea of somebody using it and it feeding to a griefbot made me feel sick to my stomach.
When ‘that sounds like them’ starts to feel like enough
One of the most revealing parts of the study was that accuracy did not always seem to be the most important thing.
Participants generally cared more about whether the chatbot felt emotionally right.
A particular term of affection could make a response feel authentic, whereas an unusually formal sentence could ruin the effect. Some replies apparently felt wrong because they were too long or too polished compared with the short messages the deceased person would once have sent.
This makes sense. People are not databases. Our sense of somebody is held through rhythms, habits, jokes, repeated phrases and tiny social cues. A chatbot does not need a complete record of a life to produce something recognisable. It may only need enough detail to imitate the shape of the relationship.
That is also where things start to become uncomfortable.
A reply can feel like Dad without containing anything Dad ever thought.
It can sound caring without caring. It can provide reassurance without knowing whether reassurance is appropriate. It can produce an apology, forgiveness or final message simply because those are the kinds of responses language models have learned to generate.
The machine has no private access to the person who died. It is not retrieving a buried truth about them. It is predicting what a suitable answer might look like from the information it has been given.
Yet emotional recognition is a powerful thing. It means that when a sentence lands correctly, it can feel discovered and deeply meaningful rather than simply generated.
We are all familiar with the strange jolt that accompanies hearing someone reproduce a dead relative’s mannerism or expression, or seeing a young relative who has grown up to look like them. For a moment, something of the person seems to return.
AI can manufacture that jolt repeatedly and on demand.
The risk is not always that someone will believe the dead person has literally returned but that the difference stops feeling important.
The ghost is partly made by the grieving person
In the study, participants sometimes treated the more distant chatbot as though it were the deceased person anyway.
Even when the system spoke about Gran in the third person, users addressed it as Gran. They asked direct questions and pulled the exchange towards a conversation with the dead, even though the researchers had designed that version to behave more like a memorial guide.
That tells us something important about how these encounters work.
The machine generates the words, but the grieving person helps turn them into something meaningful.
The user supplies memories and context, recognises certain details and ignores others. They explain away small mistakes and might ask questions that guide the conversation into emotionally meaningful territory. When something feels right, they give it weight.
This doesn’t mean they are deliberately deceiving themselves, but that meaning is being built between the person and the system.
From a skeptical point of view, that process is extremely familiar, right?
A reading delivered by a psychic medium is rarely just a stream of precise information delivered to a silent sitter. The sitter responds. They confirm, reject, reinterpret, and search their memories. A vague statement may become specific once it’s linked to a person or event.
A mistake during a psychic reading can be repaired during the conversation. The statement may be linked to another relative, treated as symbolic rather than literal, or reinterpreted in light of information supplied by the sitter. The final sense of accuracy is often produced through the whole conversation, not contained in the original words alone.
Something similar can happen with an AI ghost.
The model generates a plausible statement. The grieving person supplies the significance.
This doesn’t make the resulting emotion any less real, but it does mean we need to be thoughtful about where we believe the meaning comes from.
A digital medium without supernatural claims
A griefbot is not a spirit medium in the traditional sense. It doesn’t claim to contact an afterlife, and most users will understand that the reply is being generated by software.
But it can still perform some of the same human role; it gives the dead a voice.
The griefbot answers questions that can no longer be answered by the deceased. It creates the feeling of an ongoing exchange, and offers words that can be understood as approval, forgiveness, comfort, or even advice.
In some ways, this could make the technology more persuasive than conventional mediumship. The output is personalised using genuine information about the person who died. The creators may train the AI on real messages, recordings, or videos, so it can reproduce familiar language rather than offering general statements from a stranger in a séance room.
There is no need to believe in spirits. The system only needs to sound right often enough.
'A reply can feel like Dad without containing anything Dad ever thought.
For skeptics, this creates an awkward but important challenge. It is not enough to say, “The chatbot is not really your loved one.”
Most users already know that.
The better questions are about what the interaction encourages people to do with that knowledge.
Can someone knowingly enter the illusion, take comfort from it, and then leave without harm? Probably, in some cases.
Could repeated use begin to alter their memories of the person? Possibly.
What happens when the bot gives advice the deceased person would never have given? What happens when it invents a confession, an apology, or an opinion about a living relative? Who is responsible when a generated answer becomes emotionally important?
And what happens when the system is owned by a company whose interests go beyond helping someone through their grief?
A memorial photograph does not send notifications encouraging you to sustain the relationship. A grave does not offer a paid subscription tier (though give it time…). A saved voicemail does not learn which responses keep you engaged for longer.
Once grief becomes a product, the experience designed to comfort and the experience designed to profit may start to look uncomfortably similar.
What are we actually creating?
Some users could treat an AI memorial as a structured way to revisit memories, much as they might use a journal or speak to an empty chair during therapy. Others may find it intrusive, upsetting, or fundamentally wrong. Neither response is irrational.
The problem begins when a moving or comforting experience is presented as evidence that the technology is helpful for grief, or when comfort is mistaken for accuracy. A chatbot can produce a phrase that feels exactly like something your mother would say, but that does not mean it came from your mother, reflects her wishes, or carries her authority.
Instead, the system created a convincing answer, and you recognised something meaningful in it.
That recognition may be touching and helpful in the moment, but it can also make us less likely to notice how much of the encounter we are constructing ourselves. The same thing can happen during a reading with a psychic medium, where meaning is often built between the statements offered and the sitter’s own memories and interpretations.
None of this makes the study worthless. In fact, the researchers deserve credit for looking at how people actually respond to these systems instead of treating the subject as a distant thought experiment.
The participants’ reactions should also not be dismissed too quickly. Some of them clearly found the conversations moving, and it would be arrogant to tell them that their comfort was meaningless simply because it came from software.
But the study does not establish that griefbots are beneficial treatments. It doesn’t tell us what happens after months or years of use, or how someone in acute grief might respond, or what could happen without a researcher quietly filtering the machine’s mistakes.
It shows that, under controlled conditions, a small group of willing participants could find a brief conversation with an AI recreation emotionally powerful.
That alone deserves our attention.
Not because the technology has solved grief, or brought anybody back, but because it has shown how little is needed to create the feeling of return.
A few memories, a recognisable phrase, and a system capable of predicting the next likely sentence.
And, on the other side of the screen, somebody who dearly misses them.
The Ghost Geek is an award-winning science blog about spooky things, written by Hayley Stevens, a former ghost hunter who is now studying for a Master’s in Forensic Psychology and uses scientific scepticism to investigate strange claims.