They can answer questions the deceased never answered in life.
For as long as humans have mourned, they have sought ways to keep the dead close — in photographs, in stories, in the objects left behind. Now, a cluster of technology startups is extending that ancient impulse into new and unsettling territory, offering AI-powered simulations of deceased loved ones trained on the digital traces a life leaves behind. At monthly subscription prices, families can now hold something resembling a conversation with the dead — a development that forces a reckoning not only with grief, but with what we owe the departed and what we risk asking of ourselves.
- Startups like Séance AI and re;memory are already selling monthly subscriptions that let grieving families see, hear, and converse with AI-generated simulations of deceased loved ones built from photos, voice recordings, and social media data.
- Researchers found these 'generative ghosts' can produce eerily convincing responses from surprisingly little personal information — a few sentences about someone's personality fed into a chatbot can yield something that sounds remarkably like them.
- Users are acutely sensitive to false notes: a single phrase the deceased would never have used is enough to shatter the illusion and make people want to disengage entirely.
- Developers draw a firm line between their tools and deepfakes, arguing the intent here is comfort rather than deception — but the field itself acknowledges it has barely begun to ask whether digital resurrection helps people grieve or quietly prevents them from doing so.
Jack Manning's sister died more than a decade ago, and what stayed with him wasn't only the grief — it was the exhausting labor of honoring her memory through bake sales and fundraising runs. The experience left him wondering why grief should require such effort, and why technology couldn't help carry some of the weight.
Now a doctoral candidate at the University of Colorado Boulder, Manning studies what the field calls 'generative ghosts' — AI chatbots trained on the digital remnants of a person's life: emails, voice recordings, social media posts, video clips. Companies like Séance AI, You Only Virtual, and re;memory are already selling access to these simulations, with animated avatars that speak in the deceased's actual voice available for as little as $19.99 a month.
In a paper published this past June, Manning and his colleagues documented how people actually respond to these tools. Participants provided brief personal details about lost loved ones — memories, personality traits, characteristic phrases — and then interacted with an AI trained on that information. Many were startled by how convincing the results felt, even when built from very little data.
The research revealed that authenticity is everything. Users strongly preferred interacting with the AI in the first person, as though the deceased were speaking directly to them. And they were quick to recoil when the simulation rang false — a single misplaced term of endearment was enough to make the whole experience feel like a betrayal rather than a comfort.
Manning's colleague Jed Brubaker is careful to distinguish these tools from deepfakes, noting that the intent here is solace, not deception. But the harder questions remain largely unanswered: whether these simulations help the living process loss, or whether they create a new kind of dependency — a digital relationship with someone who is no longer there to grow or change or say goodbye.
Jack Manning's sister died more than a decade ago. What he remembers most vividly is not the loss itself, but the grinding, years-long effort to honor her memory in a way that felt true. Bake sales. 5K runs. Hundreds of hours of fundraising to build something meaningful. It was exhausting work, and it got him thinking: why should grief require such labor? Why couldn't technology help?
Today, Manning is a doctoral candidate in information science at the University of Colorado Boulder, and he's studying a technology that might have answered his younger self's question. It's called a generative ghost—an AI chatbot trained on the digital traces a person leaves behind: emails, photos, voice recordings, social media posts, video clips. Feed the algorithm enough of these fragments, and it can construct a simulation that talks back. A version of the dead that the living can still have a conversation with.
The sector is young but growing. Companies like Séance AI, You Only Virtual, and re;memory are already selling access to these digital resurrections. Re;memory charges $24 a month to create three custom avatars from photographs and voice recordings. Séance AI offers animated versions—avatars that can smile, tilt their head, speak in the deceased's actual voice—for $19.99 monthly. The pitch is straightforward: instead of simply remembering your parent, you can see them and hear them as they were, moving and speaking in real time.
In a June paper published by the Association for Computing Machinery, Manning and his colleagues, including associate professor Jed Brubaker, documented how people actually respond to these digital ghosts. Study participants submitted simple surveys about their lost loved ones—positive memories, childhood details, core personality traits. Then they interacted with an AI trained on that information. The result surprised many of them. "Participants were fascinated by how much an LLM can do with so little information," Manning explained. A few sentences about your grandfather's manner of speaking, fed into a free ChatGPT account, can produce something that sounds remarkably like him.
There are gradations of sophistication here. The simplest version—call it a death bot—merely plays back recordings verbatim, like a voice memo from beyond. It cannot improvise, cannot respond to new questions, cannot simulate the texture of how someone actually communicated. The more advanced generative ghosts go much further. They can generate novel responses. They can answer questions the deceased never answered in life. They can sustain a conversation that feels, to the person on the other end, like talking to someone who is still here.
But authenticity matters enormously. The researchers found that users strongly preferred interacting with the AI in the first person—as if the chatbot were the deceased person themselves, speaking directly—rather than as a third-person avatar narrating about them. And they were ruthlessly sensitive to inauthenticity. If the bot used a term of endearment—"champ," "pal"—that the actual person would never have said, users wanted to stop talking. The simulation had to feel true, or it felt like a betrayal.
There is a natural comparison to deepfakes, those fabricated videos designed to deceive the public by putting words in a public figure's mouth. But Brubaker draws a sharp distinction. "The fundamental premise of a deepfake is the intent to deceive," he said. "The intent of generative ghosts is not to deceive another person." These are tools built for grief, not for fraud. They exist in the space between memorial and conversation, between honoring the dead and talking to them. Whether that space is healthy, whether it helps people process loss or delays it, whether it offers solace or creates a kind of digital dependency—those are the questions the field is only beginning to ask.
Citas Notables
Participants were fascinated by how much an LLM can do with so little information.— Jack Manning, doctoral candidate in information science
The fundamental premise of a deepfake is the intent to deceive. The intent of generative ghosts is not to deceive another person.— Jed Brubaker, associate professor at University of Colorado Boulder
La Conversación del Hearth Otra perspectiva de la historia
What made you decide to study this instead of, say, just building the product yourself?
I watched my family spend years raising money for a physical memorial. It was meaningful, but it was also exhausting. I wondered if there was a way to honor someone that didn't require that much friction. But I also realized I didn't understand what this technology actually does to people emotionally. So I wanted to study it first.
When someone talks to one of these AI ghosts, do they know they're talking to a machine?
Yes, they know. But that doesn't seem to matter as much as you'd think. What matters is whether it sounds like the person. If it uses their phrases, their way of thinking, their voice—then the interaction feels real in a way that matters to the person grieving.
That sounds like it could be dangerous. What if someone becomes dependent on talking to the ghost instead of processing the grief?
That's the question we're asking now. We don't have long-term data yet. But I will say this: the people we studied were very discerning. They rejected inauthenticity immediately. They seemed to know the difference between a memorial and a replacement.
Do you think your sister would have wanted this?
I think she would have found it strange at first. But I also think she would have appreciated that someone could talk to her, ask her questions, hear her voice again. Whether that's healthy or not—I'm still trying to figure that out.
What happens to these ghosts when the company goes out of business?
That's a problem we haven't solved. Right now, if a startup fails, the digital version of your loved one disappears. There's no legal framework for digital inheritance yet. It's one of the things we need to think about as this grows.