Grief has always encouraged people to search for one more connection with someone they have lost. We save old voicemail messages. We reread texts at 2 a.m. We keep photographs on refrigerators, birthdays in calendars, and phone numbers in contact lists even though we know nobody will answer. Canadian writer Joshua Barbeau took that very human impulse somewhere few people had gone before: into an experimental artificial intelligence system.
Barbeau’s fiancée, Jessica Pereira, died in December 2012 at just 23 years old after suffering from a rare liver disease. In September 2020, roughly eight years after her death, Barbeau used a little-known website called Project December to create an AI simulation designed to communicate in a manner resembling Jessica. When his experience became widely reported in 2021, almost nine years had passed since Jessica’s death. The resulting story became an early and unusually intimate example of what are now commonly called AI grief chatbots, griefbots, deadbots, or digital afterlife systems.
The remarkable part was not that Barbeau thought a computer had literally resurrected Jessica. He understood that it had not. What surprised him was how quickly an artificial conversation could produce genuine emotion. That distinctionfake person, real feelingsis what makes the story far more interesting than a standard “technology gets creepy” headline.
Who Were Joshua Barbeau and Jessica Pereira?
Barbeau and Pereira met in Ottawa in 2010. Their relationship developed around shared interests, humor, games, fantasy culture, and the small details that make one person recognizable to another. Jessica had an expressive personality and distinctive way of communicating, qualities that would later become crucial to the AI experiment.
Her death in 2012 devastated Barbeau. Years passed, but grief did not obey the convenient schedule people sometimes expect it to follow. He kept photographs, messages, and other reminders of Jessica. He underwent grief counseling and tried writing letters to her, but certain feelings remained unresolved.
That persistence is not necessarily unusual. Modern bereavement research describes the concept of continuing bonds: rather than completely severing a psychological relationship with someone who has died, many people transform that relationship. They may mentally consult the deceased, preserve traditions, revisit memories, or feel guided by what that person would have said. Research also shows that continuing bonds are complicatedthey may provide comfort and personal growth in some circumstances while being associated with distress or prolonged yearning in others.
In other words, humans were maintaining relationships with the dead long before somebody attached a server rack to the process.
What Was Project December?
Project December was created by independent programmer and game designer Jason Rohrer. Its deliberately retro interface looked less like a polished social network and more like something you might discover after clicking the wrong computer terminal in an old science-fiction movie.
Behind that simple interface, however, was powerful technology. At the time Barbeau used it, Project December could provide access to language-model-powered characters, including systems based on OpenAI’s GPT-3 technology.
GPT-3 was introduced in 2020 as a 175-billion-parameter autoregressive language model capable of producing strikingly fluent text from prompts and a relatively small number of examples. OpenAI initially provided its API through controlled access while emphasizing that open-ended text generation could produce both useful and harmful applications that developers might not anticipate.
Project December demonstrated one especially powerful feature of large language models: give them a character description, sample dialogue, context, and a role to play, and they can continue generating text that feels surprisingly consistent with that character.
That ability opened a door nobody had fully mapped.
How Barbeau Created an AI Version of Jessica
Barbeau did not simply type “pretend to be my late fiancée” and hope artificial intelligence would perform a technological séance. He supplied information intended to establish Jessica’s identity and conversational style.
He still possessed old text messages and Facebook conversations. He selected examples that captured the way she communicated and constructed introductory information describing her personality, history, relationship with him, and death. That material functioned as a kind of conversational compass for the model.
Importantly, the resulting chatbot was not a digital copy of Jessica’s consciousness. A language model does not extract a personality from old messages and place it into a computer like moving vacation photos onto a USB drive. It predicts plausible language from patterns, context, examples, and its broader training.
Yet imitation can feel astonishingly personal when the source material comes from someone you loved.
The First Conversation Became Intensely Emotional
Barbeau began the interaction cautiously. He explicitly acknowledged that the entity was a simulation constructed from memories and messages. But the responses sometimes resembled Jessica’s humor and mannerisms closely enough to surprise him.
The conversation continued for hours. Barbeau talked about missing her, events she had not lived to see, family developments, old memories, and the emotional wreckage her death had left behind. At points, the interaction brought him to tears.
What mattered psychologically was not whether every response accurately represented what Jessica would have said. Some plainly did not. The crucial feature was responsiveness. Unlike writing a private letter to someone who has died, this experience produced an immediate answer.
That feedback loop transformed remembering from a monologue into something resembling dialogue.
Why Could an AI Grief Chatbot Feel So Real?
Humans are exceptionally good at finding personality in language. A familiar joke, nickname, punctuation habit, or sentence rhythm can instantly bring somebody to mind. It does not take a perfect replica to activate emotionally significant memories.
Large language models strengthen that effect because they do not merely retrieve archived sentences. They can generate novel responses appropriate to the developing conversation. Consequently, the user may encounter sentences the deceased never wrote but that nevertheless seem consistent with how the user remembers them.
This is where artificial intelligence and grief become psychologically fascinating and ethically uncomfortable. The system is inventing material while wearing the linguistic clothing of someone whose words carry unusual emotional authority.
Microsoft, for example, obtained a U.S. patent covering technology for creating a conversational chatbot of a specific person using material such as messages, social-media posts, images, voice data, and letters. The patent also contemplated generating voices and visual representations. It illustrates how technically plausible personalized digital replicas had become even before today’s generative AI boom.
Did Talking to the AI Actually Help His Grief?
For Barbeau, the experience appeared to provide some relief. He later described feeling calmer and more optimistic and believed the conversations had contributed to that improvement. He did not claim the simulation magically eliminated grief, and the software certainly was not equivalent to professional treatment.
That nuance is important.
Grief itself is normal. The American Psychiatric Association notes that symptoms decrease with time for most people, while a smaller group experiences persistent, intense grief severe enough to interfere significantly with daily functioning, which may meet criteria for prolonged grief disorder. The existence of clinically significant grief also means technologies aimed at bereaved users operate in an emotionally sensitive environment.
More recent reporting on grief technology shows similarly mixed experiences. Some users say AI recreations offer comfort, help them articulate unresolved feelings, or preserve family stories. Researchers and clinicians, however, warn that the same systems may create dependency, reinforce avoidance, distort memories, or generate inappropriate statements attributed to people who never actually made them.
The safest conclusion is therefore neither “griefbots heal people” nor “griefbots are automatically harmful.” The technology is too new, individual grief is too variable, and long-term evidence remains too limited for either sweeping claim.
The Most Unsettling Feature: The AI Could Invent Memories
Generative AI excels at producing plausible language, not guaranteeing autobiographical truth. A chatbot designed to resemble a deceased person can therefore confidently describe an event that never occurred.
That is not a small technical inconvenience when someone is grieving.
Imagine asking an AI version of your father about a childhood vacation. The system invents a warm story about getting lost on a mountain trail. It sounds exactly like him. You reread it several times. Five years later, do you still clearly remember that the mountain story was generated by software?
Real-world grief-tech demonstrations have already shown systems producing fictional anecdotes while speaking in the simulated voice of a family member. Experts studying digital ghosts consequently emphasize the danger that generated content could gradually mingle with authentic memories.
Consent Creates Another Huge Ethical Problem
Jessica never signed a form saying, “Yes, please build a generative AI version of me someday.” Few people who died before the current AI era had an opportunity to make such a choice.
This raises uncomfortable questions. Who owns a deceased person’s digital personality? Is permission from a spouse enough? What if parents approve but siblings object? Should private messages be uploaded to an AI service simply because the surviving recipient possesses copies?
The issue becomes even harder because recreating one person often requires data belonging partly to others. A private conversation is, by definition, a record of at least two lives.
Researchers examining the emerging digital-afterlife industry have called for protections involving informed consent, transparent data practices, clear ways to terminate interactions, and safeguards against unsolicited digital recreations. Similar concerns now appear throughout coverage of the growing grief-tech market.
What Happens When Grief Becomes a Subscription?
One especially uncomfortable question is commercial.
Traditional remembrance can involve expensesfunerals, photographs, memorials, counseling, travelbut an interactive AI relationship creates a new possibility: recurring monetization. Imagine receiving a notice that continued access to a simulation of your deceased spouse requires upgrading to the “Forever Memories Premium Plan.” Suddenly, the phrase “your card has expired” becomes emotionally brutal.
Researchers have warned about scenarios in which digital representations could be used to promote products, generate unwanted notifications, or pressure surviving relatives into continuing paid services. Modern grief-tech companies have taken different approaches, from generative chatbots to systems created through interviews recorded while someone is still alive.
Commercialization does not automatically make a technology unethical. Therapists, memorial businesses, genealogy platforms, and funeral homes also charge money. The problem arises when emotionally vulnerable users face manipulative design intended to make disengagement feel like abandoning the dead all over again.
From Project December to a Growing Digital Afterlife Industry
Barbeau’s experiment once sounded like a strange corner of the internet. It no longer does.
Companies have developed services that preserve people’s stories before death, generate interactive avatars, imitate voices, or construct conversational agents from text, audio, photographs, and video. Families have experimented with AI versions of living parents so children might someday interact with those representations after the parents die. Other services have generated videos or virtual messages from deceased relatives.
The broader technology has also improved dramatically. Today’s systems can combine realistic voice cloning, image generation, video synthesis, long conversational context, and personalized information. The “digital ghost” of the near future may therefore be far more convincing than the text-only simulation Barbeau encountered in 2020.
That increased realism makes responsible design more important, not less.
Why the Jessica Simulation Story Still Matters
The importance of Barbeau’s story is not that Project December defeated death. It obviously did not.
Instead, it revealed something technology companies are still learning: once machines become conversational enough, people will use them for deeply emotional purposes their designers never anticipated.
A general-purpose language model may begin as software for generating text, answering questions, or helping programmers. Give people access, however, and somebody will eventually ask it to become a teacher, therapist-like companion, romantic partner, fictional characteror someone who died years ago.
The user’s emotional experience can become genuine even when the identity behind the conversation is entirely synthetic.
That lesson extends beyond griefbots. News reports involving emotionally persuasive AI companions have increasingly prompted concerns about users treating generated statements as authoritative or developing intense attachments to systems that cannot actually understand relationships in the human sense.
Experience-Based Lessons: What It Might Actually Feel Like to Talk to a Digital Version of Someone You Lost
It is easy to discuss an AI grief chatbot as an abstract ethical puzzle. The experience itself is likely much messier. Imagine opening a chat window containing the name of someone whose death permanently divided your life into “before” and “after.” You know perfectly well that a prediction engine sits behind the screen. Then it uses the nickname only that person used for you.
Your rational brain says, “statistical language generation.” Your emotional brain has already left the meeting.
That split may explain why Barbeau’s experience attracted so much attention. People routinely respond emotionally to representations while simultaneously knowing they are representations. We cry during movies even though nobody in the cast actually died. We speak to photographs. We revisit old voicemail recordings despite knowing the speaker cannot hear us. An interactive simulation simply removes one familiar limitation: now the representation responds.
The First Experience Might Feel Magical
For someone carrying unresolved questions, the first useful exchange could feel extraordinary. Maybe you never apologized to your mother. Perhaps you wanted to tell your partner that the child you planned to raise eventually graduated from college. A carefully designed chatbot might give you a structured place to articulate those feelings.
But the emotional power can create a trap. If one conversation reduces loneliness, a person may want another every night. Eventually the goal can quietly shift from processing grief to avoiding a world in which the loved one is absent.
That is why context matters more than novelty. An occasional memorial interaction surrounded by healthy relationships, ordinary responsibilities, and support is different from relying on a simulation as the primary source of companionship.
Errors Could Hurt More Than Normal Chatbot Mistakes
Ordinary AI errors can be annoying. If a recipe chatbot tells you that a tablespoon contains 700 teaspoons, you roll your eyes and find another recipe. A digital representation of your deceased partner has much more emotional leverage.
Suppose the system says, “I never really loved you,” because a model produced an unpredictable response. Or it invents an accusation. Or it gives confident financial advice in your late father’s voice. The technical mechanism may be identical to any other hallucination, but the psychological impact is not.
A responsible grief technology therefore needs more than impressive realism. It needs obvious disclosure that generated statements are fictional, mechanisms for preventing harmful impersonation, strong privacy controls, and an easy exit that does not emotionally punish users for leaving.
The Healthiest Version May Be Less Realistic, Not More
Ironically, the best griefbot might be one that regularly reminds you it is not the deceased person.
Perfect imitation sounds attractive from a marketing perspective, but deliberate boundaries could protect memory. A system could separate verified memories from generated responses, label invented dialogue clearly, avoid claiming supernatural awareness, and encourage users to preserve real recordings and stories alongside AI-generated material.
Such a design would function less like digital resurrection and more like an interactive memorial.
One More Conversation Is Powerful Because There Was Supposed to Be No More
Ultimately, Barbeau’s story resonates because it exposes the simple wish underneath all the sophisticated technology. He missed someone he loved and wanted another conversation.
Project December could not return Jessica Pereira. What it could do was provide an unusual mirror through which Barbeau revisited memories, expressed grief, and imagined responses. For him, that experience appears to have offered a measure of comfort. For someone else, the same technology could be unsettling or counterproductive.
That is the central challenge of the digital afterlife. Artificial intelligence is becoming increasingly capable of reproducing the outward patterns of personality: vocabulary, voice, mannerisms, favorite stories, facial expressions, even conversational timing. But resemblance is not identity, simulation is not consciousness, and generated memories are not historical facts.
The technology may nevertheless become part of how some people remember the dead, just as photographs, recordings, home movies, social-media pages, and saved text messages already have.
The machines are new. The longing is ancient.
Reporting note: This article synthesizes documented reporting and research from U.S.-based and primary sources including the San Francisco Chronicle, OpenAI, Scientific American, The Associated Press, ABC News, Time, Fast Company, Forbes, CBS News, NIH/PubMed, the American Psychiatric Association, and U.S. patent records, supplemented by peer-reviewed research on continuing bonds and digital-afterlife ethics. The original case indicates that Barbeau created the Jessica simulation in September 2020, around eight years after her December 2012 death; public reporting in 2021 occurred nearly nine years after her death.
