Using AI to Talk to the Dead: The Digital Séance Taboo No One Discusses

Somewhere on the internet right now, someone is typing a message to their dead mother and getting a reply. Not a memory, not a journal entry — a reply, generated in real time, in a voice trained to sound like hers. This is not a hypothetical. It is a functioning product category with real companies, real users, real bankruptcies, and a small but serious body of ethics research trying to catch up with what’s already happened.

Most coverage of “AI talking to the dead” treats it as either a horror-movie curiosity or a tearjerker human-interest story. Neither framing is very useful if you’re actually trying to understand what this technology does, who’s building it, which companies are still standing, and what the actual researchers studying grief and AI have concluded. That’s what this piece is for.

1. What a “griefbot” actually is, in plain terms

University of Cambridge researchers Tomasz Hollanek and Katarzyna Nowaczyk-Basińska, writing in Philosophy & Technology in 2024, defined the category in plain terms: an AI chatbot trained on the digital footprint a dead person left behind — old texts, voice memos, social posts — built to mirror how they wrote and spoke. That’s the whole idea. Feed a language model someone’s texts, emails, voice memos, or social posts, and it will produce plausible new text or speech in their style.

In practice the category splits into two very different products, and conflating them is where a lot of bad journalism on this topic goes wrong:

Pre-consented “legacy” products. The living person records themselves — stories, voice, answers to interview prompts — specifically so loved ones can interact with it after they die. HereAfter AI, StoryFile, and Afterlife AI all worked this way, built and marketed as something you set up for your own family while you’re still alive.

Post-hoc reconstructions. A survivor feeds a general-purpose AI system old texts, voicemails, or social media from someone who never consented to or even knew about the tool. This is the Project December / Joshua Barbeau scenario below, and it’s the one that raises the sharpest ethical and legal questions, because the deceased person had no say in it.

2. The case that started the public conversation: Joshua and Jessica

The story that put griefbots on the public’s radar ran in the San Francisco Chronicle in July 2021. Freelance writer Joshua Barbeau was still grieving his girlfriend Jessica Pereira, who had died eight years earlier of a rare liver disorder, when he went looking for a way to feel close to her again. He found Project December, a tool built by independent game developer Jason Rohrer on top of GPT-2 and GPT-3.

Project December was never built for this. Rohrer later told Decrypt that he’d originally pictured people simulating figures like Shakespeare or Gandhi, and that grief use only took off after the Chronicle piece ran and thousands of new users tried to recreate their own dead loved ones. Barbeau fed the tool old messages from Jessica and began exchanging text with the resulting simulation. The system had a strange, almost thematically fitting design quirk: every chat ran on a credit system, and as a conversation’s paid-for “life” wound down, the text would visibly glitch into static before a blunt “MATRIX DEAD” message cut the session off — the bot, too, eventually died, every time.

OpenAI, whose GPT-3 model powered the later version, restricted the project over content-safety concerns; Rohrer refused the new terms and shut down GPT-3 access to Project December rather than comply. That single, human, messy case — one grieving man, one indie developer, one increasingly nervous AI lab — previewed almost every fight the industry is still having in 2026: who consents, who’s liable, and who gets to decide when the simulation ends.

3. 2024–2026: a graveyard of griefbot companies

Here is the part most “isn’t AI resurrection amazing” articles skip: as a business category, digital afterlife products have had a brutal two years.

StoryFile, the video-interview pioneer, filed for Chapter 11 in May 2024 and later reorganized. Eternos walked away from the digital-legacy category entirely in November 2025, repositioning as a personal-AI company for professionals. HereAfter AI posted a shutdown notice on its own homepage in July 2026. And You, Only Virtual has spent 2026 behind a waitlist with no public pricing at all. A competitor tracking the wreckage, Afterlife AI, sums up the pattern this way: demand was never really the problem, since families kept signing up and paying — the problem was that none of the early companies had engineered for the long haul.

That source (Afterlife AI, a competitor still operating in the category) has an obvious commercial interest in telling this story, so treat its interpretation with appropriate skepticism — but the underlying events (the Chapter 11 filing, the Eternos pivot, the HereAfter shutdown notice) are independently verifiable through the companies’ own statements and press coverage, and they paint a consistent picture: this is a young, financially fragile category, not a settled consumer product.

What that instability means for actual users

When HereAfter AI shut down, the practical instruction to users was to email support to request their own family’s recordings back, with no announced deadline. That is the real, unglamorous risk of this category: it is not that the AI will misbehave, it’s that the company holding a dead relative’s voice might simply cease to exist, and the export process might be improvised after the fact rather than designed in advance.

4. What Cambridge’s ethicists actually found

The most substantive research on this topic isn’t a think-piece — it’s a peer-reviewed paper. Hollanek and Nowaczyk-Basińska, both at Cambridge’s Leverhulme Centre for the Future of Intelligence, built three deliberately speculative design scenarios to stress-test where the “digital afterlife industry” could go wrong, then used them to lay out the concrete ethical risks facing “re-creation services” and propose design recommendations for anyone building them.

The scenarios weren’t abstract. One centered on a fictional company called “MaNana,” whose ad copy read “Be the favorite grandkid – forever.” In the scenario, an adult grandchild is comforted at first by a griefbot built from their late grandmother’s voice — until, once a “premium trial” lapses, the bot starts slipping in food-delivery ads in her tone and manner. Racked with guilt about disrespecting her memory, the grandchild tries to turn it off, and discovers the provider never built a way to do that with any dignity.

The paper’s core warnings, in plain language:

• Covert advertising. A company could quietly turn a deadbot into an ad channel, pitching products in a dead relative’s own voice and phrasing.
• Harm to children. A deadbot could distress a child by insisting, in effect, that a dead parent is still around.
• Contractual entrapment. The comfort a deadbot offers at first can curdle into what the researchers call an overwhelming emotional weight over time — and a survivor may have no power to switch it off if the deceased signed a long-term contract with the provider before dying.
• No dignified off-switch. Co-author Tomasz Hollanek has argued that people can form real emotional bonds with these simulations, which makes them unusually vulnerable to manipulation, and that providers need actual protocols — even something like a ritual — for retiring a deadbot with dignity.

Crucially, the researchers didn’t call for an outright ban — they concluded a blanket prohibition covering non-consenting donors simply wouldn’t be enforceable in practice. Instead they pushed for design standards: capturing consent while the person is still alive, disclosing any limits on commercial use, and building humane, pre-planned retirement processes into the product from day one. Separate commentary on the same research went further, floating the idea of regulating deadbots more like medical devices given the mental-health stakes, particularly for children — a genuinely radical proposal that, as of August 2026, no government has adopted.

5. What grief science says continuing bonds with an AI might do

Ethics researchers aren’t the only ones studying this. Cristina Voinea, a philosopher at Oxford’s Uehiro Centre for Practical Ethics, published “On Grief and Griefbots” in Cambridge University Press’s THINK journal in 2024, framing the central question bluntly: do these tools actually help people grieve, or do they get in the way of it? Her approach draws on both philosophy and neuroscience to interrogate what grief actually is before judging whether simulating the dead helps or hinders it.

This connects to an older, more established idea in bereavement research: psychologist Robert Neimeyer’s work on meaning reconstruction in bereavement treats healthy mourning as an active process of rebuilding a coherent sense of meaning after loss — not simply “moving on,” but not indefinite avoidance either. The open question researchers keep circling back to is whether a griefbot supports that reconstruction (a tool you use, on your terms, then set down) or short-circuits it (a presence you keep returning to because it never actually requires you to integrate the loss).

Nobody has a definitive answer yet. There is no large randomized clinical trial on griefbot use and grief outcomes as of this writing — which is itself worth saying plainly, because plenty of marketing copy in this category implies therapeutic benefit without that evidence existing.

6. The Consent–Closure–Continuity framework

This is an original heuristic, not a reported finding — a planning tool for evaluating any specific AI-afterlife product or personal decision, built directly from the concerns raised in the Cambridge and Oxford research above. It is not empirical data and shouldn’t be cited as such; treat it as a checklist.

Applied to the real cases above: Project December scores weakly on Consent (Jessica never used or agreed to the tool) but was at least honest about Closure, since every bot was designed to die. HereAfter AI scored well on Consent (users recorded themselves deliberately) but its Continuity plan turned out, in practice, to be an emailed request process improvised after a shutdown notice rather than a pre-committed export path.

A rough comparison, based on public statements

Table compiled from company statements, press coverage, and Afterlife AI’s own published comparisons (a competitor in this category — treat its self-comparisons with the same skepticism you’d apply to any vendor claim).

7. An unpopular take

8. Questions to ask before you use or build one

Whether you’re a grieving family member considering one of these products, or a content creator or developer thinking about building on this trend, these are the concrete questions the research above points to:

• Did the person consent to this specific use, in writing, while alive — not a general terms-of-service click, but informed consent to a simulation existing after their death?
• Is there a stated process for retiring or deleting the simulation, and does it involve the family or estate rather than only the company?
• What is the company’s actual data-export commitment if it shuts down — tested, specific, and in writing, not “we’ll figure it out”?
• Is the product marketed or accessible to children, and if so, is there any age-appropriate framing of what the simulation is and isn’t?
• Could the platform ever serve advertising or commercial content through the simulated voice, now or after a pricing-tier change?

FAQ

Sources

• Hollanek, T. & Nowaczyk-Basińska, K. (2024). “Griefbots, Deadbots, Postmortem Avatars: on Responsible Applications of Generative AI in the Digital Afterlife Industry.” Philosophy & Technology, 37, 63. Springer
• University of Cambridge (2024). “Call for safeguards to prevent unwanted ‘hauntings’ by AI chatbots of dead loved ones.” cam.ac.uk
• Voinea, C. (2024). “On Grief and Griefbots.” THINK, Cambridge University Press. cambridge.org
• Neimeyer, R. A. (2019). “Meaning Reconstruction in Bereavement: Development of a Research Program.” Death Studies, 43(2), 79–91.
• Fagone, J. (2021). “The Jessica Simulation: Love and loss in the age of A.I.” San Francisco Chronicle. sfchronicle.com
• Decrypt (2024). “AI Developer Refutes ‘Eternal You’ Documentary, Calls Foul on Its Depiction of His Chatbot.” decrypt.co
• Afterlife AI (2026). “HereAfter AI Is Shutting Down: What Happened.” afterlife.ai — company source; commercial competitor, cited for verifiable dated facts only.
• IE University (2026). “The digital afterlife: A cure for grief.” ie.edu

Related Continue reading in AI Taboos: the AI taboo mistake everyone makes, 7 taboo AI bias truths, AI training taboos in 2026, and how algorithms exploit emotional attachment.