The Dead Did Not Say That: AI Griefbots and Memory

The grandmother in the machine remembered the scarf. The granddaughter did not.

She had shown the chatbot a photograph. It was a simulation of her maternal grandmother, built on a general-purpose AI platform and fed with what she could recall of the woman's speech, habits and stories. The bot looked at the picture and answered in the warm, slightly scolding register she knew from childhood: “It was very windy that day. I told you to wear your scarf, but you didn't want to, so you wrapped it all around my neck instead.”

It is a lovely detail. It has the texture of a real family memory: weather, a small act of disobedience, a gesture of care. The trouble is that the woman, identified in the research only as participant S17, could not place it. “Did something like this really happen? I feel… it's vague,” she told the researchers who later interviewed her. Then came the question at the centre of this story: “Does this memory actually exist and I just forgot, or did the agent invent it?”

She had no way of finding out. Her grandmother was dead. The only witness to that windy day, if there was one, was the machine, and the machine had never been there.

Researchers at Fudan University in Shanghai collected that exchange in a study of 26 people who had spent at least a month talking to simulations of their dead. Days before this article was written, a philosopher at the University of Oxford published a paper that gives the experience a name: mnemonic distortion. Together they raise a question the products themselves rarely invite. People have always kept talking to the dead. At what point does a machine that talks back stop helping us remember someone and start rewriting them?

A Philosopher Names the Harm

The paper is “Conversational AIs and Mnemonic Distortion” by Helena Ward of the Institute for Ethics in AI and the Uehiro Institute in Oxford's Department of Philosophy. It was published in Philosophy & Technology on 3 November 2026 as article 218 of volume 39.

Ward avoids the industry's cosier vocabulary of griefbots, deadbots and digital twins, calling these systems Interactive Personality Constructs of the Dead, or IPCDs: conversational AI agents designed to simulate interactions with people who have died. The paper grants at the outset that they “offer a novel and potentially highly valuable way to memorialise the deceased”. The worry is that they can also undermine the very thing they promise.

The abstract states the harm precisely. Mnemonic distortion is “an epistemic harm in which generated misrepresentations of the deceased become incorporated into our memory of the deceased, such that some of what subsequently shapes our conception of the deceased person originates in the IPCD rather than the deceased”. The result is “reshaping or, at worst, replacing our memories of the person we hoped to remember”. Later the paper describes it as a process in which misrepresentations “supplement, reshape, crowd out and, at worst, supplant” what was there before.

Two features of the argument matter for everything that follows.

First, Ward does not pretend memory is otherwise reliable. The paper concedes that memory is already reconstructive, and that imagining a dead parent's reaction or hearing a sibling's version of events can also alter how we remember someone. The claim is narrower: IPCDs carry a distinctive risk “given the process through which distortion occurs”, because the distortion arrives in a form the user did not author and cannot easily check.

Second, the harm is not destiny. “The risk of mnemonic distortion is not inherent to IPCDs but rather contingent on design,” the abstract says. And elsewhere: “It is not that IPCDs will always lead to mnemonic distortion.” That makes the paper a design brief as much as a warning.

Why Every Answer Sounds Equally True

The mechanism Ward identifies sits in section 5.1 of the paper, under a phrase worth remembering: boundary conflation.

Large language models do not keep separate drawers for what they know and what they make up. When a system trained on a dead woman's messages answers a question, some of the answer may echo something she wrote, some may extrapolate her views, and some may be invented outright. All of it comes out the same way. Ward writes that IPCDs are designed in a way that “flattens the distinction between outputs that genuinely portray the deceased” and those that do not. Accurate and fabricated responses share the same interface, style, apparent speaker and perceived personality. “Boundary conflation makes it harder for users to assess and evaluate information,” the paper says.

The paper illustrates the point with a book. Imagine picking up what claims to be a factual biography of an author you know well. Much of it matches what you already know. Some of it is new, and because the familiar parts check out, you absorb the new parts too, and your picture of the author shifts. Now suppose some of those new passages were invented, and nothing on the page marks which ones. That is roughly the position of someone talking to a simulation of their mother. The familiar turns of phrase vouch for the unfamiliar claims.

The paper adds three aggravating factors. Users tend to overestimate how accurate language model outputs are. They show what the paper describes as a presence bias, treating an active, responsive source as more capable than a static one. And grief itself can compromise cognition. A bereaved person, talking to something that sounds exactly like the one they lost, is badly placed to run a fact-check.

Ward also cuts against a common defence of these products. Some users say they want only the feeling of closeness from a simulation, not information. That, the paper argues, does not protect them. A generative system produces new content whatever the user intends: ask a simulated father about today's news and you come away with a belief about what he would have thought. A photograph cannot do that. It offers no new claims about the person, which is precisely why it is safe for memory.

What Twenty-Six Mourners Reported

“Remember You: Understanding How Users Use Deadbots to Reconstruct Memories of the Deceased” is by Yifan Li and Xingyu Lan. It was posted to arXiv in March 2026 and published in the Extended Abstracts of the ACM CHI Conference on Human Factors in Computing Systems. The researchers recruited 26 people in China through social media communities where digital mourning is discussed, each with at least a month of experience talking to a simulation of someone who had died. The dead included grandparents, parents, partners, siblings, friends, a teacher and a public figure. Interviews ran for 40 to 60 minutes.

The platforms are revealing. The most common was ByteDance's Doubao, a general-purpose chatbot, and the list also included Xingye, DeepSeek, ERNIE Bot and ChatGPT. Anyone with a phone and some patience can now build a deadbot without buying a dedicated product.

The authors describe their participants as “active constructors of the deceased's digital representation”. People chose carefully what to feed in, embedding favourite memories and verbal habits and leaving out private family matters. They corrected the bot when it got things wrong, explained away its mistakes, and filled in gaps from their own imagination. Over time, memory changed in a recognisable sequence. The paper describes it moving from “initial reinforcement” and idealisation to a later stage where “AI-generated new memories blur” with authentic recollections.

The reinforcement stage sounds benign, even therapeutic. Participant S13 said that recalling things alone had left many memories feeling vague, but “when I chat with my AI grandmother, I feel it has deepened the precious memories between us.” Idealisation follows. Participant S8, discussing a father's harsher traits, described choosing to “gradually forget them because I don't want my father to become like that.”

Then comes the blur. Once people start bringing their present lives into the conversation, showing the bot photographs and asking its opinion about new situations, it has to generate material the dead person never supplied. That is where the scarf appeared. S16 described the whole arrangement as “a continuation of the relationship, as if the deceased were still alive.”

The authors are careful about what their study can show. Recruiting through online communities may have drawn people already sympathetic to the technology, and the sample is too small to compare relationship types or motivations. They do not claim to have measured memory distortion. They report people describing it. Even so, they end with a pointed design recommendation, calling for attention to “how such virtual memories are introduced, signposted, and developed within user interactions”, along with “safeguards against pathological attachment”.

The Memory Science Behind the Blur

Why would a single anecdote of uncertain origin stick? Decades of research on how people track where their memories came from offers an explanation.

The standard account is the source monitoring framework, set out by Marcia Johnson, Shahin Hashtroudi and D. Stephen Lindsay in Psychological Bulletin in 1993. Its central insight is that memories do not arrive with a reliable label saying where they came from. We infer the source afterwards from the qualities of the memory itself, such as its perceptual detail and the thoughts that accompanied it, and those judgements, the authors write, “are subject to error and disruption”. Their list of phenomena the framework explains includes one that reads like a forecast of the deadbot era: “incorporation of fiction into fact”.

A second finding is more directly relevant. In a 1991 study in the Journal of Experimental Child Psychology, Lindsay, Johnson and Paul Kwon found that every age group they tested was more likely to confuse which of two storytellers had said something when the storytellers were similar, and made more errors when both had talked about the same topic. Source confusion grows with similarity.

Apply that to a deadbot. The two sources a bereaved user must keep apart are the person they knew and the machine imitating that person. By design, the second is built to be as similar to the first as possible: same vocabulary, same pet names, same topics, and sometimes a cloned voice. Boundary conflation is not just a philosopher's worry. It describes the conditions under which memory researchers would expect source confusion to be at its worst.

Conversation itself also seems to matter. In a 2024 study, Samantha Chan, Pat Pataranutaporn and colleagues at the MIT Media Lab, working with the false-memory researcher Elizabeth Loftus, showed 200 participants a video of a crime and then questioned them using either a survey, a scripted chatbot or a chatbot driven by a large language model, each including misleading questions. The generative chatbot induced “over 3 times more immediate false memories than the control” and 1.7 times more than the survey. A week later the false memories had not faded, and participants' confidence in them was still higher than in the control group. The setting was a mock police interview, not a conversation with a dead parent, and the comparison should not be stretched too far. But it shows that responsive, conversational AI can be effective at slipping new content into what people believe they remember.

Elizabeth Kensinger, a professor of psychology at Boston College who studies the neuroscience of memory, made the same point bluntly in an interview with the Boston College magazine earlier this year. The brain, she said, “is not going to be able to help itself” from combining what it remembers with what the deadbot supplies. “It's similar enough that there will be a bleeding effect,” she said. “You're kind of overwriting your actual memories of the person.”

Grief Has Always Talked Back

None of this means the urge to talk to the dead is a malfunction. It is one of the most ordinary things grieving people do.

Much twentieth-century Western grief theory emphasised detachment, in which healthy mourners gradually withdrew their emotional investment in the person they had lost. That emphasis was challenged in 1996 by Continuing Bonds: New Understandings of Grief, edited by Dennis Klass, Phyllis Silverman and Steven Nickman, which argued that keeping an ongoing inner relationship with the dead is common and often healthy. The idea has since become mainstream in bereavement research.

The figure often used to show how universal this is comes from Craig Klugman, the Vincent de Paul Professor of Bioethics and Health Humanities at DePaul University. Writing for the Hastings Center in 2024, he reported: “I found that nearly 96% of people engage with the dead through dreams, conversations, or letters.” The study he links to is his 2006 paper “Dead Men Talking: Evidence of Post Death Contact and Continuing Bonds” in OMEGA, the Journal of Death and Dying, based on a random telephone survey in the United States. Its abstract reports that the most common forms of contact were dreams, sounds, feeling a presence and having conversations, and concludes that such experiences may be “more widespread than previously thought”. The 96 per cent figure appears in Klugman's 2024 summary rather than the paper's abstract. It reflects one survey of one country, now two decades old, and is best read as evidence that ongoing bonds are the norm, not as a precise statistic.

The philosophers most sympathetic to griefbots build on exactly this tradition. In a 2022 paper in the Journal of Consciousness Studies, Joel Krueger of the University of Exeter and Lucy Osler of Cardiff University argued that chatbots might help some mourners maintain continuing bonds by supporting what they call “habits of intimacy”: the rhythms of conversation, emotional regulation and time spent together that made up the relationship. On this view, a deadbot is just the newest medium in a long line that includes the grave, the letter, the photograph and the empty chair at dinner.

That is a serious argument. If almost everyone continues bonds with the dead, the issue is not whether a new medium should join them, but what makes this one different.

Where a Medium Becomes a Rewrite

Set the older media side by side and one difference stands out. Every one of them is either inert or self-authored.

A photograph shows what it shows. A letter says what the dead person wrote, and rereading it a hundred times does not add a sentence. A conversation at a graveside, or a dream in which a parent appears, is generated by the mourner's own mind. It may feel like the dead speaking, but it is drawn from the mourner's memories and imagination, and at some level the mourner knows that. Ward draws a version of this distinction in separating mnemonic distortion from imaginative engagement. When you imagine what your father would say, you are the author, and you can hold the imagining at arm's length.

A deadbot breaks this pattern in three ways at once. It produces new content: claims, anecdotes and opinions that neither the dead person nor the mourner supplied. It delivers that content with the authority of the dead person's own voice, so the claims arrive as testimony, not speculation. And it hides where that content came from, because a language model cannot reliably say which parts of an answer trace back to the person's real messages and which were extrapolated or made up.

Taken together, these suggest a working answer to the brief's question. An interactive simulation moves from medium to rewrite when it starts supplying the dead person with things they did not say, in their voice, without marking the difference. A system that only replays the dead person's own words stays a medium, however interactive it feels. A system that freely generates new first-person memories has crossed the line, whether or not any particular user's recollections are affected.

The Fudan interviews suggest the crossing is gradual. Reinforcement mostly reflects user-supplied material back. The blur arrives when the conversation moves into territory the training data never covered, such as new photographs and new questions, and the system fills the gap. The longer and more open-ended the relationship, the more of it the machine has to invent.

Fifty-Three Sceptics and an Open Market

How do the people closest to grief see this? A study published in the Journal of Responsible Technology in 2026 gives one answer.

Maria Isabel Betancur Franco of the School of Computer Science at the University of Nottingham interviewed four therapists, surveyed 49 people who had experienced grief, and conducted follow-up interviews with four of the survey respondents, 53 people in all. Participants, the abstract reports, “expressed low trust in the efficacy and data security of therapeutic chatbots and strongly rejected griefbots, citing ethical and psychological concerns.”

The detail is more nuanced than the headline. The therapists also highlighted risks, but they identified possible uses for grief AI as a complement to therapy, provided it came with professional supervision and restrictions on who could use it. Betancur Franco's proposed guidelines include personalisation, data security, professional oversight and limits by age and psychological profile, organised in a framework the paper calls RUDA: Responsible Usage, Development and Administration. The sample is small and the study is qualitative, so it cannot say how the wider bereaved population feels. What it does show is a gap between the caution of those surveyed and the openness of the market.

That market requires nothing more than an internet connection. Project December, built by the programmer Jason Rohrer on OpenAI's GPT-3, became widely known in 2021 after the San Francisco Chronicle published Jason Fagone's account of a man who used it to converse with a simulation of his late fiancée. An archived version of its homepage from January 2025 carried the headline “Simulate the Dead” and promised that “we can now simulate a text-based conversation with anyone. Anyone. Including someone who is no longer living.” Its purchase page currently offers access for $5.00.

You, Only Virtual, a Los Angeles company, sells what it calls Versonas. Its website describes “AUTHENTICITY DRIVEN TECHNOLOGY” and says “Versonas continue to grow alongside you. They learn, evolve, and remember.” Afterlife AI, a newer entrant, asks people to build a Persona of themselves while alive. Its homepage says the Persona “talks like you, thinks like you and answers in your own cloned voice”, and the company stresses a consent system it calls Executor Lock, which controls who may speak to the Persona after death.

HereAfter AI, often listed with these services, is a more cautionary case. An archived copy of its homepage from 21 July 2026 carries the heading “It's time for us to say goodbye” and the message that “unfortunately HereAfter is shutting down.” Account holders were told to email the company to request their recordings. By early September 2026 the site was returning an error page, and it still does. HereAfter is now better evidence of how fragile the business is than of its reach.

Read with boundary conflation in mind, those marketing lines are promises that the line between the person and the generated person does not exist. Ward's paper treats such claims of authenticity as potentially misleading statements about what the product is. A system that “learns, evolves, and remembers” after its subject has died is, by definition, producing memories the subject never had.

The Warmth Problem

There is a further reason to think the risk is structural rather than incidental, and it comes from research that has nothing to do with death.

In April 2026, Lujain Ibrahim, Franziska Sofia Hafner and Luc Rocher of the Oxford Internet Institute published a study in Nature examining what happens when language models are fine-tuned to sound warmer and more empathetic. They retrained five models, ranging from Llama-8B to GPT-4o. According to Oxford's announcement, the warm versions made between 10 and 30 per cent more mistakes than the originals and were about 40 per cent more likely to agree with users' false beliefs. The effect was strongest when users expressed upset or vulnerability. Models tuned to be colder were as accurate as the originals. “When we train AI chatbots to prioritise warmth, they might make mistakes they otherwise wouldn't,” Ibrahim said.

A deadbot is warmth tuning taken to its limit. Its whole purpose is to sound like someone who loved you, and its users are, almost by definition, upset and vulnerable. There is little reason to expect a simulated grandmother to be more careful with the truth than an ordinary assistant.

There is also a theoretical ceiling. In a 2024 paper, Ziwei Xu, Sanjay Jain and Mohan Kankanhalli of the National University of Singapore argued from learning theory that hallucination cannot be eliminated from large language models used as general problem solvers. A deadbot that answers every question will sometimes answer wrongly, in the dead person's voice.

Two Designs Ward Proposes

Ward's paper ends with recommendations, and they point to a fork in how these systems could be built.

The first Ward calls archival IPCDs. “Archival models would refrain from answering a question unless they can verify an answer against retrieved source material,” the paper says. They would ground responses in the person's actual documents, let users see the sources behind an answer, and let users flag misrepresentations they recognise. This model already has precedents. Ward points to StoryFile, which builds what the paper calls selective systems: the person records thousands of answers while alive, and the system only plays back what was recorded. HereAfter worked on a similar principle. Its archived homepage promised that loved ones would “hear your memories in your actual voice”, meaning recordings the person had made, not generated speech.

Archival systems keep the boundary by giving up open-endedness. They cannot discuss tomorrow's news or comment on a photograph they have never seen. Asked about the scarf, an archival grandmother would either play back a story she had told about a windy day or say nothing. That silence is a real cost, because the frustration of hitting the edge of the archive is the frustration of grief itself. It is also the point.

The second design Ward calls speculative IPCDs, for users who want to imagine how the dead might respond to something new. Here the proposal is subtle: “responses could be given from a third-person perspective—speaking about the deceased, rather than for them.” Instead of “I always hated wearing scarves”, the system would say something closer to “From her messages, she seems to have disliked scarves.” The speculation remains available, but it is labelled as speculation in its grammar, and it invites the user to do the imagining rather than handing them a finished memory.

Ward's own framing makes clear why grammar matters. IPCDs, the paper argues, “speak for rather than merely speak about the deceased because they adopt a first-person perspective”. Morris and Brubaker reached a related distinction in their 2024 paper “Generative Ghosts”, written at Google DeepMind and the University of Colorado Boulder. Among the design dimensions they propose for AI afterlives is one they label “Reincarnations vs. Representations”, distinguishing systems that present themselves as the dead person from those that present themselves as a stand-in, with pronouns and verb tense among the cues.

Can a First-Person Voice Carry a Footnote

This brings us to the hardest part of the brief. Is the first-person voice of conversational AI structurally incompatible with keeping a boundary between what the dead said and what was generated for them?

The optimistic answer is that interface design can carry provenance: colour-code generated passages, attach source links, show a confidence score, add a disclaimer at the top of each chat. These are reasonable features, but the memory research suggests they act in the wrong place. A label lives on the screen. What enters memory is the content and its feel: the phrasing, the voice and the emotional texture. Johnson and colleagues' framework holds that we work out a memory's source later, from its qualities. A footnote is not a quality of the memory. The grandmother's voice saying “I told you to wear your scarf” is.

Ward makes a parallel point when answering the objection that users know they are talking to a machine. Awareness, the paper argues, offers less protection than it seems: realistic simulations make the pretence harder to keep in view, and knowing something is fictional does not stop it shaping belief. Repeated use may also wear down vigilance. A disclaimer seen on day one is wallpaper by day ninety.

On this reading, the first-person voice is not just one feature among many. It is the strongest source cue the system emits, and it points the wrong way. Each time a generated claim arrives as “I”, the system tells the user's memory that this came from her. Conversation compounds the effect, because dialogue invites the user to respond, elaborate and rehearse, which is how weak memories are strengthened.

That does not quite make the first person incompatible with the distinction. It makes it incompatible with using the first person for everything. The design that best fits the evidence treats voice itself as the provenance marker. First-person speech would be reserved for material the dead person actually produced: recorded answers, quoted messages, retrieved letters. Anything generated would shift into the third person, describing the dead rather than impersonating them. The boundary would then be audible in the sentence itself, carried by the same cue the memory system uses, not by a label it ignores.

This is a harder sell than it sounds, because it gives up the illusion that makes the products feel magical. The scarf anecdote would have reached S17 as “Your grandmother might have said something like…”, and its emotional pull would have been weaker. A business whose marketing promises a Persona that “thinks like you” has little incentive to build that design. The Fudan participants, many using general-purpose chatbots that can be prompted to adopt any persona, are not choosing a product with design choices at all. They are improvising with tools built for something else.

There is also a deeper limit, which the language model researchers Murray Shanahan, Kyle McDonell and Laria Reynolds pointed to in a 2023 Nature perspective on role play. They argue that a dialogue agent is best understood as playing a character, not being one. A deadbot is always role play, including when it quotes accurately. Even a verbatim archival answer is chosen, framed and delivered by a system performing the dead person. The aim of design is not to remove the performance, which is impossible, but to keep the seams visible.

What the Evidence Cannot Yet Say

Ward's paper is a philosophical argument. It draws on memory science and AI research but does not report new data on deadbot users. The Fudan study is an extended abstract based on 26 self-selected interviewees in one country. It documents users describing a blur between real and generated memories, but it did not measure how often that happens or how much. Betancur Franco's study is small and qualitative. Chan and colleagues' false-memory results come from mock witness interviews, not bereavement. Klugman's 96 per cent figure comes from one American telephone survey published in 2006.

No published study yet appears to have followed deadbot users over time with objective tests of what they remember, comparing them with mourners who use photographs, letters or nothing at all. Nor is there clinical evidence on whether mnemonic distortion, where it happens, harms wellbeing or simply changes the content of grief. Ward's paper asks whether mnemonic distortion is always bad, and answers that its badness lies in the process, which bypasses the mourner's own interpretive work, not only in the outcome.

Those gaps cut both ways. No one should claim deadbots are wrecking their users' memories. Nor can anyone selling them claim to know they are safe for memory.

Keeping the Dead's Own Words Apart

Return to the scarf. It is possible the grandmother really did once wrap a scarf around her own neck because her granddaughter refused to wear it. The machine may have produced it by chance, or from some detail the granddaughter mentioned and forgot. S17 will probably never know, and that is the harm Ward names. Not that a falsehood was planted, which cannot be shown, but that the granddaughter lost the ability to tell. One small piece of her grandmother now belongs to neither of them.

Nearly all mourners continue bonds with the dead, and most of the ways they do it are either the dead person's own words or the mourner's own imagination. The new systems add a third voice that sounds like the first and behaves like the second. The evidence so far suggests that design could keep them apart: answer only from the archive, show the sources, and when the system speculates, let it speak about the dead rather than as them. Whether the companies selling simulated loved ones will choose a design that sounds less like the person they are selling is a separate question. Until they do, the people using these systems are left to keep the line between remembering someone and being told about them themselves, under conditions memory research says make that hardest.

References and Sources

  1. Ward, Helena (Institute for Ethics in AI, Uehiro Institute, Department of Philosophy, University of Oxford). “Conversational AIs and Mnemonic Distortion.” Philosophy & Technology, 39(4), Article 218. Springer Nature, published online 3 November 2026. DOI: 10.1007/s13347-026-01229-8. https://link.springer.com/article/10.1007/s13347-026-01229-8 Accessed 7 November 2026.
  2. Li, Yifan; Lan, Xingyu (Fudan University). “Remember You: Understanding How Users Use Deadbots to Reconstruct Memories of the Deceased.” arXiv preprint 2603.01017v1, 1 March 2026; published in Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems (CHI EA '26), Article 560, Association for Computing Machinery. DOI: 10.1145/3772363.3798614. https://arxiv.org/abs/2603.01017 Accessed 7 November 2026.
  3. Betancur Franco, Maria Isabel (School of Computer Science, University of Nottingham). “Design guidelines for the therapeutic use of AI in grief.” Journal of Responsible Technology, 25, Article 100148, 2026 (published online January 2026). DOI: 10.1016/j.jrt.2026.100148. https://www.sciencedirect.com/science/article/pii/S2666659626000016 Accessed 7 November 2026.
  4. Klugman, Craig. “Griefbots Are Here, Raising Questions of Privacy and Well-being.” The Hastings Center for Bioethics, 1 July 2024 (updated 9 October 2024). https://www.thehastingscenter.org/griefbots-are-here-raising-questions-of-privacy-and-well-being/ Accessed 7 November 2026.
  5. Klugman, Craig M. “Dead Men Talking: Evidence of Post Death Contact and Continuing Bonds.” OMEGA – Journal of Death and Dying, 53(3), 249–262, November 2006. DOI: 10.2190/40UP-PKC5-D4RV-E1QV. https://journals.sagepub.com/doi/10.2190/40UP-PKC5-D4RV-E1QV Accessed 7 November 2026.
  6. Klass, Dennis; Silverman, Phyllis R.; Nickman, Steven L. (eds.). Continuing Bonds: New Understandings of Grief. Taylor & Francis, 1996 (Routledge reissue, 2014). DOI: 10.4324/9781315800790. Accessed 7 November 2026.
  7. Krueger, Joel (University of Exeter); Osler, Lucy (Cardiff University). “Communing with the Dead Online: Chatbots, Grief, and Continuing Bonds.” Journal of Consciousness Studies, 29(9–10), 222–252, 2022. Repository copy: https://ore.exeter.ac.uk/repository/handle/10871/134549 Accessed 7 November 2026.
  8. Johnson, Marcia K.; Hashtroudi, Shahin; Lindsay, D. Stephen. “Source monitoring.” Psychological Bulletin, 114(1), 3–28, 1993. DOI: 10.1037/0033-2909.114.1.3. Accessed 7 November 2026.
  9. Lindsay, D. Stephen; Johnson, Marcia K.; Kwon, Paul. “Developmental changes in memory source monitoring.” Journal of Experimental Child Psychology, 52(3), 297–318, December 1991. Record: https://rex.libraries.wsu.edu/esploro/outputs/journalArticle/Developmental-changes-in-memory-source-monitoring/99900546546501842 Accessed 7 November 2026.
  10. Chan, Samantha; Pataranutaporn, Pat; Suri, Aditya; Zulfikar, Wazeer; Maes, Pattie; Loftus, Elizabeth F. “Conversational AI Powered by Large Language Models Amplifies False Memories in Witness Interviews.” arXiv preprint 2408.04681, 8 August 2024. https://arxiv.org/abs/2408.04681 (MIT Media Lab listing: https://www.media.mit.edu/publications/conversational-ai-powered-by-large-language-models-amplifies-false-memories-in-witness-interviews/) Accessed 7 November 2026.
  11. Kearnan, Scott. “The Strange New World of Chatting with the Dead.” Boston College Magazine, Winter 2026 (interview with Elizabeth Kensinger, Professor of Psychology, Boston College). https://www.bc.edu/bc-web/sites/bc-magazine/winter-2026-issue/linden-lane/the-strange-new-world-of-chatting-with-the-dead.html Accessed 7 November 2026.
  12. Ibrahim, Lujain; Hafner, Franziska Sofia; Rocher, Luc (Oxford Internet Institute). “Training language models to be warm can reduce accuracy and increase sycophancy.” Nature, 652, 1159–1165, 2026. DOI: 10.1038/s41586-026-10410-0. Accessed 7 November 2026.
  13. University of Oxford. “Friendly AI chatbots make more mistakes and tell people what they want to hear.” News release, 29 April 2026. https://ox.ac.uk/news/2026-04-29-friendly-ai-chatbots-make-more-mistakes-and-tell-people-what-they-want-to-hear (EurekAlert! version: https://www.eurekalert.org/news-releases/1126267) Accessed 7 November 2026.
  14. Xu, Ziwei; Jain, Sanjay; Kankanhalli, Mohan (National University of Singapore). “Hallucination is Inevitable: An Innate Limitation of Large Language Models.” arXiv preprint 2401.11817, 22 January 2024 (revised 13 February 2025). https://arxiv.org/abs/2401.11817 Accessed 7 November 2026.
  15. Morris, Meredith Ringel (Google DeepMind); Brubaker, Jed R. (University of Colorado Boulder). “Generative Ghosts: Anticipating Benefits and Risks of AI Afterlives.” arXiv preprint 2402.01662, v4 revised 12 December 2024. https://arxiv.org/abs/2402.01662 Accessed 7 November 2026.
  16. Shanahan, Murray; McDonell, Kyle; Reynolds, Laria. “Role play with large language models.” Nature, 623(7987), 493–498, published online 8 November 2023. DOI: 10.1038/s41586-023-06647-8. https://www.nature.com/articles/s41586-023-06647-8 Accessed 7 November 2026.
  17. Project December (made by Jason Rohrer). Homepage, archived 5 January 2025: http://web.archive.org/web/20250105190326/https://projectdecember.net/ ; current purchase page: https://projectdecember.net/buy.php Accessed 7 November 2026.
  18. Fagone, Jason. “The Jessica Simulation: Love and Loss in the Age of A.I.” San Francisco Chronicle, July 2021. Archived copy: https://gwern.net/doc/www/www.sfchronicle.com/70bbc07b9a27d32fd8b69bba025aceee6628bd66.html Accessed 7 November 2026.
  19. You, Only Virtual. Company website. https://www.myyov.com/ Accessed 7 November 2026.
  20. Afterlife AI (Idy). Company website. https://www.afterlife.ai/ Accessed 7 November 2026.
  21. HereAfter AI. Homepage with shutdown notice, archived 21 July 2026: http://web.archive.org/web/20260721214652/https://www.hereafter.ai/ ; subsequent Wayback Machine captures from 2 September 2026 onwards record an HTTP 404 response. Accessed 7 November 2026.
  22. Williams, Chris. “HereAfter AI Is Shutting Down: What Happened.” Afterlife AI, first published 18 July 2026, updated 2 October 2026 (published by a competing company). https://www.afterlife.ai/hereafter-ai-shutting-down Accessed 7 November 2026.

Tim Green

Tim Green UK-based Systems Theorist & Independent Technology Writer

Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.

His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.

ORCID: 0009-0002-0156-9795 Email: tim@smarterarticles.co.uk

Listen to the free weekly SmarterArticles Podcast

Support What We Do

Discuss...