Griefbots and the Perils of Digital Immortality

New artificial intelligence tools can make chatbots that are scarily good at mimicking dead loved ones. Should they exist?

Thank you for reading The Garden of Forking Paths. I *love* writing for you, but I rely on reader support to keep it sustainable. Some of this article is for paid subscribers only. Please consider upgrading (for just $4/month) to access it—and 150 other essays—in full. You can also support my work by buying my new book: FLUKE.

---

I: Mind the Gap: The Ghost at Embankment

In late 2012, shortly before Christmas, an old woman sat down in London’s Embankment station. Unlike the other passengers, she wasn’t there to travel. While everyone else ignored the pre-recorded safety announcement, she was eagerly awaiting it. When it came, she was dismayed. Then, she began to cry.

As Tube trains whizzed past, commuters swirling around her, Margaret McCollum wiped tears from her eyes. It had been five years since her husband, Oswald Laurence, had died. But this was the day Margaret felt like he was truly gone forever.

When staff at the station came over to comfort Margaret, they asked her what was wrong. What she told them moved them to tears, too.

She explained that her late husband, Oswald, was an actor. In the late 1960s, he had recorded the “Mind the Gap” announcement for the Northern Line on London’s sprawling underground network. Even after he died in 2007, McCollum could always come to the station, sit on the platform, and listen as her husband’s distinctive, sonorous voice boomed a safety warning to passengers, reminding them about the gap between the train doorway and the platform. It was a ritualized comfort, a tiny audio link to the man she loved and deeply missed.

This time, though, Oswald’s voice had disappeared, replaced in a fresh digital upgrade of the Tube announcements. McCollum’s visceral connection with her husband was gone.

After hearing her story, unsung heroes at Transport for London decided to track down the old recordings of Oswald. With considerable effort, they managed to digitize his voice and use it for one station—Margaret’s station—the northbound platform of Embankment Tube.¹ It is still used there today, the only outlier on the entire underground network.

II: Keeping the Dead Alive, Somehow

For much of human history, the death of a loved one meant they could only be immortalized through mental representations, or physical objects that they had either made or touched. What they looked or sounded like, the contours of their face, the swirls of their wrinkles, or how they would behave in a fresh situation were all relegated to memory and imagination.

There has, however, long been a widespread human impulse to cling to living remnants of the dead, prolonging our attachment to them.

Perhaps the most extreme example comes from the Toraja people, an isolated group in the lush, remote expanse of Indonesia’s Sulawesi island. Death for the Toraja is gradual, not abrupt, as the family lives alongside the deceased body—sometimes for years—while they save up money for the extravagant funeral. In local belief, until that funeral has taken place, the corpse hasn’t yet truly died, so the family cares for the body, keeping them as part of their daily life. Some families share portions of every meal in offerings to the body. The living may even chat to the dead, a son telling the corpse of his Mom about his day, about his dreams, about his impending marriage. For the Toraja, grief is gradual, slow, elongated.

Closer to our experience, in modern Western culture, new technology has allowed more vivid physical reminders of the dead. Death masks were taken to allow people to see a visible likeness, as plaster fades slower than memory. In Victorian Britain, the advent of photography gave rise to death photos, post-mortem photographs which may strike us as macabre, but served as a lifeline to people for whom capturing visual memories was a new-fangled luxury. And, though it’s often forgotten in his biography, Thomas Edison tried to create a “spirit phone” to communicate with the dead.

In more recent times, the accessibility of recording technology allowed some people who knew they were going to die—terminal patients with long, slow deaths, for example—to choose to shoot a video or record an audio message for their young children to watch later in life. The line between life and death inexorably blurred ever so slightly with every new technology, but it was still unambiguous: it was always clear that a recording was a one-way transmission produced before death, not a revived representation of someone who had died.

Now, for the first time in history, that has changed. The line between life and death is being blurred in ways that were previously impossible.

III: “Jessica” and the Disturbing Magic of Griefbots

In late 2012, a young Canadian woman named Jessica Pereira was admitted to hospital. She had severe complications from a liver transplant that she had received at the age of nine. Now 23 years-old, her liver was beginning to fail.

Her boyfriend, Joshua Barbeau, was completely smitten with Jessica. A social misfit on the autism spectrum, Barbeau had found someone who understood him. With medical tubes down Jessica’s throat, Barbeau proposed. When she was released from hospital, he promised, they would get married.

But Jessica never was released. Her condition unexpectedly worsened. On December 11, 2012, Jessica Pereira died. She was just 23 years-old.

For the next eight years—as Jason Fagone of the San Francisco Chronicle reported in a brilliantly moving feature article—Joshua couldn’t move on. For months after she died, he barely spoke to others, only able to process the grief by talking to his Border Collie. Joshua was anxious, rarely leaving the house. He had little social contact. He tried going to grief therapy, but as Fagone notes, “Most of the others in the room were in their 60s or 70s and were dealing with the loss of a life partner. Joshua was 26.”

No matter what he did, his mind kept returning to one topic: Jessica. Grief consumed him. He couldn’t let go.

Joshua was, however, adept with computers. During the early stages of the pandemic, he began to dabble in a fledgling new technology—chatbots. He was intrigued by these faux personas, including one modelled after William Shakespeare, responding to text prompts from users, all in the style of the bard.

But what caught Joshua’s attention one night was something altogether more unbelievable: the ability to shape a chatbot to mimic someone specific. All you needed was a large bank of text—such as all your texting and e-mail chat history—produced by that person, so the model could be updated to learn their style, their turns of phrase, the way they used language and emojis and punctuation.

As Fagone writes:

As Joshua continued to experiment, he realized there was no rule preventing him from simulating real people. What would happen, he wondered, if he tried to create a chatbot version of his dead fiancée?...

…That night in September, Joshua hadn’t actually expected it to work. Jessica was so special, so distinct; a chatbot could never replicate her voice, he assumed. Still, he was curious to see what would happen. And he missed her.

It worked. The digital resurrection was impressive, even convincing. Joshua began to spend more time chatting to “Jessica.” But he knew that the software he was using had a finite limit; he had paid for enough credits to spend hours upon hours with this quasi-digital reincarnation of his fiancée, but not enough for it to last forever.

And yet, as he typed, and the chatbot typed back, he was repeatedly struck by how uncanny the resemblance could be with her personality. Over and over—even though he knew this was merely software—he found himself thinking: Jessica would have said exactly that.

During his grieving process, Joshua had tried to move on, even to date someone new. With the new girlfriend, he was upfront with her about his simmering grief. He talked endlessly about memories of life with Jessica. At first, the new girlfriend found it sweet, a sign of a man who was devoted and loyal. But that devotion, that inability to let go never waned. Eventually, she couldn’t handle dating someone who was still mentally with someone else—and not just an ex, but someone who had died. They split up, and Joshua was back to where he began: alone with his grief.

When Joshua began speaking to “Jessica,” he decided to update her on his life, not only including his attempts to remember her daily, but also recounting his failed bout of trying to be with someone else. This screenshot of that chat is from the San Francisco Chronicle feature on their story:

Disturbing as this is, Joshua and Jessica’s story is—at least in part—a story about a griefbot that helped mend a broken heart. Over time, Joshua did find that he was returning to the bot less and less frequently. Rather than let his credits run out and have the digital mimic of his love “die,” Joshua left it with an ambiguous goodbye. These were their final messages:

Joshua: I’ll never stop loving you for as long as I live, and hopefully ever after. Xoxo <3 Goodnight.

Jessica: Goodnight. I love you.

IV: The Dubious Ethics of Digital Immortality

These griefbots feel different in important ways from the more ordinary practice that some grieving people endure, in which they cling to photographs, letters, mementos, or get lost for hours re-reading old messages that were sent during life.

“Jessica” was an early incarnation of chatbot technology shaped for a specific individual; Fagone’s reporting was published in July 2021. In terms of artificial intelligence, that gap of three years between then and now is akin to moving from the Stone Age to the Industrial Revolution in a blink; large language models, or LLMs, have evolved relentlessly month by month. They are now far more powerful, far more nuanced linguistic engines.

In other words, a version of “Jessica” produced today would stir much more frequent uncanny moments for Joshua, where the echoes of Jessica’s personality feel both more familiar and more natural. Even to the most disbelieving skeptic, it would be hard to shake those emotional moments of disbelief and doubt, in which users must consciously remind themselves: remember: this isn’t real.

Companies are already capitalizing on this technology. If you visit the website of Replika, the welcome page promises “the AI companion who cares: always here to listen and talk. Always on your side.” Hereafter AI allows people to upload texts, photos, videos, and audio of someone who has died. If you prefer, you can do it yourself before you die, so that your loved ones can “see your photos and hear your memories in your actual voice.” It promises that “your voice brings your personality to life so that children, grandchildren, and beyond can know what you’re really like.” The chatbot then can speak in your actual voice.

These companies are but the tip of the technological iceberg. We haven’t even begun to realize the potential to create convincing likenesses of our loved ones. But like so much of that new realm of artificial intelligence, it’s completely unregulated—and there are persuasive reasons to worry that Joshua’s experience with “Jessica” is likely to be the exception, and that most interactions involving griefbots will end in disaster.

The risks are enormous, and they raise a series of fascinating, but unnerving, ethical conundrums that demand urgent government intervention.