The Morbid Option Is the One Where You Do Nothing

Look at what is in the drawer. A will is money, executors, witnesses and signatures. A funeral plan is a venue, a coffin, music and payment terms. Life insurance is a health questionnaire, a premium and a list of exclusions. An advance directive is treatment preferences and one nominated person, for a day when you cannot speak.

Four items, four kinds of administration, and every one of them asks you to picture the room with you absent from the room. That is why the drawer feels the way it does. You are not avoiding paperwork. You are avoiding a rehearsal.

The fifth item is a conversation about the forty or sixty years you have already had: a bad manager in 1998, a house you could not afford, the holiday where everything went wrong and became the story everybody still tells. None of that asks you to imagine your own funeral, and neither does the interview.

Psychiatry stopped treating the impulse to go back over your own life as a symptom more than sixty years ago. Robert Butler named the life review in 1963, calling the increased reminiscence of older people a naturally occurring developmental process rather than something to be managed, and connecting it to late-life depression in some people and to candour, serenity and wisdom in others (Butler, Psychiatry, 1963).

Then it was measured. A meta-analysis of 128 controlled trials of reminiscence interventions found moderate improvements in ego-integrity (g = 0.64) and depression (g = 0.57), with smaller improvements in purpose in life (g = 0.48) and in death preparation (g = 0.40) (Pinquart & Forstmeier, Aging & Mental Health, 2012).

That last figure inverts the objection. Structured talking about your life improved how prepared people were for death, as a side effect of talking about something else. The task you filed next to the funeral plan does the funeral plan's emotional work better, without raising the subject. If the doubt underneath yours is really the feeling that you are not important enough to leave anything behind, start there instead. That objection is the parent of this one.

Underneath the morbid question sits a second one, bigger for most readers and nothing to do with death. Half of it is correct, and a page that told you otherwise would deserve less of your trust.

The strange half is the one nobody agreed to: a chatbot assembled from a dead person's messages, by somebody else, after that person lost the ability to say no or correct a word. That is the version in the television plots, and the version your brother pictures when you say "an AI version of me". Keep the unease. Aim it correctly.

The ordinary half is a living person answering questions they chose and naming who may hear the answers later. A letter written to be opened on a future birthday works that way. So does a camcorder interview with a grandparent, the kind that turns out to be the only recording of that voice anybody has. Nobody calls either of those weird.

Tomasz Hollanek and Katarzyna Nowaczyk-Basinska, both at the University of Cambridge, published a paper in Philosophy & Technology on 9 May 2024 analysing harms in what they call the digital afterlife industry. They split the problem across the data donor, whose data makes the simulation possible, the data recipient, who holds the deceased person's information, and the service interactant, the person talking to the system. Their recommendations included retirement procedures, transparency about what the system is, adult-only access, and mutual consent from both data donors and interactants (Hollanek & Nowaczyk-Basinska, Philosophy & Technology, 2024).

Mutual is the word doing the work, and a deadbot fails both halves. The donor is dead: cannot agree, cannot hear the result, cannot withdraw. That is why the result reads as something done to a person rather than by one, and why your instinct is reading a real ethical hole. The ethics of building Personas of people who have died covers that ground in full.

In your own case you are the donor. You pick the questions and what gets recorded. You name who may listen, and you can take a name off the list on a Thursday. One objection survives that: your children did not consent to being handed a version of their parent. True. The people you name can decline, nobody is enrolled automatically, and Personas nobody chooses to open simply sit unopened.

The feeling itself is neither squeamishness nor superstition. Kurt Gray and Daniel Wegner ran three experiments published in Cognition in 2012 and found that machines unnerve people when they seem to have experience, the capacity to feel and sense, rather than agency, the capacity to act. The same paper found the reverse for humans: a person who appears to lack experience reads as creepy too (Gray & Wegner, Cognition, 2012).

That predicts which bits of this bother you. A search box over your old emails bothers nobody. A folder of voice memos bothers nobody. A voice that answers in your timing, with your hesitations, and seems to mean what it says, bothers almost everybody for the first thirty seconds. Ordinary mind-perception machinery is firing at exactly the feature that makes the recording worth opening. Gray and Wegner tested mind perception, not grief, so the finding travels only so far, and what it buys is a better question: what rules sit around the feature doing the unsettling.

You are measuring yourself against an imagined average person, and social psychology has a name for getting that estimate wrong. Deborah Prentice and Dale Miller published a study in the Journal of Personality and Social Psychology in 1993 on pluralistic ignorance and alcohol use on campus: students believed they were more uncomfortable with campus drinking than the average student was, so they felt deviant against a misperception rather than reality, and that perceived deviance tracked with alienation from campus life. American students, alcohol, the early 1990s, nothing to do with AI. The mechanism transfers anyway: people misjudge what everybody else privately thinks, then feel strange against a figure they invented.

You have a friend in your head who finds this ridiculous. That friend may exist. You have not asked, and you are already paying the full social cost of a conversation that has not happened.

The death-adjacent version has been measured. In the Conversation Project's 2018 national survey, 92% of Americans said it was important to talk about their wishes for end of life care, while only 32% had actually had that conversation. In the same survey, 95% said they would be willing to talk, and more than half said they would feel a sense of relief in doing so (The Conversation Project, 2018).

Ninety-five per cent willing. Thirty-two per cent done. Whatever sits in that gap, it is not a population who find the subject unspeakable. Everybody is waiting for somebody else to start.

The dread is a measured miscalculation too. Across a series of experiments, conversations between strangers felt less awkward and produced more connectedness and happiness than participants predicted, and people especially overestimated the awkwardness of deep conversations compared with shallow ones. The reason was consistent: people expected others to care less about their disclosures than others actually did (Kardas, Kumar & Epley, Journal of Personality and Social Psychology, 2022). Anticipated worse than experienced, which is the feeling you had when you closed the tab.

The hours themselves are mundane. In our own sessions people laugh early, usually at their own younger self, and some ring a sibling afterwards to check a detail.

Humans devote 30 to 40% of speech output solely to telling other people about their own experiences, and self-disclosure activates the mesolimbic dopamine system, including the nucleus accumbens and ventral tegmental area. In the same studies, participants gave up money for the chance to disclose about themselves (Tamir & Mitchell, PNAS, 2012).

The strangeness of a new medium wears off, too. Claude Fischer's history of the telephone industry documents that the industry resisted sociable use of the phone and then, in the 1920s, shifted to endorsing sociability, because customers were already using the phone that way (Fischer, Technology and Culture, 1988). The University of California Press summary of his book puts the same point about women's use of the phone as something the industry "first vigorously discouraged and then later wholeheartedly promoted". Ringing a friend for no reason, the eventual entire point of owning a telephone, was once improper use of the device. A sense of impropriety around a new medium is frequently a norm rather than a fact, and norms move. Yours is allowed to move, and allowed not to.

Start building your legacy Free build, 50 memories, no card required.

Doing nothing moves the subject into a worse room, at a worse hour, with your answers missing.

A systematic review of 40 studies covering 2,854 surrogate decision makers, more than half of them family members, found that at least a third carried a negative emotional burden from deciding on someone else's behalf. The effects were often substantial and typically lasted months, sometimes years: stress, guilt about the decisions, doubt about whether they had been right. Knowing which treatment matched the person's preferences was frequently cited as reducing the burden (Wendler & Rid, Annals of Internal Medicine, 2011).

Saying things out loud in a durable form is not a soft intervention. In a randomised controlled trial of 309 Melbourne hospital inpatients aged 80 and over, among those who died within six months, end of life wishes were known and followed in 86% of the facilitated planning group against 30% of controls. Families of patients who died in the intervention group had significantly lower stress, anxiety and depression scores than control families (Detering et al., BMJ, 2010).

Set that beside a finding that cuts the other way, because quoting one without the other is a sales technique. A systematic review of sixteen studies covering 2,595 surrogate and patient pairs found that surrogates, whether next of kin or chosen by the patient, predicted treatment preferences with 68% accuracy, and that neither being chosen nor prior discussion improved that accuracy (Shalowitz, Garrett-Mayer & Wendler, Archives of Internal Medicine, 2006).

The difference is mechanical rather than motivational. A vague chat over Christmas lunch is not the same object as specific answers recorded in a form that survives you. Talking does not turn the people who love you into mind readers. Recorded specifics do change what happens in the room.

Morbid is not an hour spent describing your best day. Morbid is your family in a corridor years from now, guessing at what you wanted, while your answers were available for free the entire time you were alive.

The gap between wanting that conversation and holding one is not morbidity. That is procrastination wearing a serious coat.

Say "recording". Say "interview". Name who gets access. Avoid every phrase that sounds like beating death, because those are the phrases people argue with. Nobody argues with a recording. Plenty of people will argue with "a version of me that lives on", and they are arguing with the phrase rather than with what you are doing.

Skip "AI version of me" as your opener, because the listener hears a film plot before they hear the word recording. Lead with the recording and let them reach the AI part themselves, which they usually do within a sentence.

If it comes with a joke about robots, the joke is somebody buying a second to work out how they feel. Laugh, then give them the one-breath version: same technology as the creepy thing, opposite permission structure, and I am the one giving permission.

The person who reacts worst on the night is often the one who rings weeks later to ask how the recording works. Their first reaction was about somebody of their own that nobody ever recorded, and scepticism is easier to say at a table.

Point what is left at the mechanism. Who can open this, when, and who verifies them. If the answer is "anyone with the password", the consent was decorative and your discomfort is correct. If it involves evidence, a named person, a waiting period and a notification to everyone else on the list, the consent has teeth. Ask that of every service, including mine.

Opting out is not neutral. The VALL-E system described by Chengyi Wang and colleagues at Microsoft, in a paper posted in January 2023, synthesises personalised speech from a three-second enrolled recording of a speaker the model has never heard before (Wang et al., arXiv, 2023). Your voice is already in voicemails, in a wedding video, in the answerphone message you recorded in 2014 and forgot, and protecting your voice from AI cloning is a separate job. So the choice is between a version you made deliberately, with a named person at the gate, and whatever somebody assembles later without asking anyone.

The law is drawing the same line. Article 50(4) of the EU AI Act, Regulation (EU) 2024/1689, requires deployers of an AI system generating deepfake image, audio or video content to disclose that the content is artificially generated or manipulated, and under Article 113 the Regulation applies generally from 2 August 2026. One caveat, checked 30 July 2026: the European Commission's own AI Act service desk warns that the provision has been amended by the Digital Omnibus on AI Regulation proposal and that the published text has not caught up, so check the wording before relying on detail. Regulators have made disclosure a legal line as well as a moral one.

Ours works like this. A guide called Idy runs the interview, follows what you say rather than a fixed script, and returns to threads you left open. Answers are stored as memories you can read, edit or delete at any time, your voice is cloned from the recordings you make in session rather than scraped from anything online, and you name who may speak with the result. Nobody else gets in, including anyone who turns up claiming to be family.

Nothing releases while you are alive, and no setting quietly changes that default. Release happens through Executor Lock™, a seven day evidence verified process: a person you named files a request, an evidence file is required rather than an assertion accepted, you and the requester are both emailed while the window runs, you can stop the whole thing yourself, and everyone else you named hears only once the Lock opens. The Lock snapshots everything you built, exactly as it stood on the trigger day. Deleting the account removes your account row and the stored audio.

Now the tradeoffs. Release is deliberately slow, so nobody gets access on the day they ask, and the people waiting pay that cost. The voice clone is only as good as the audio you give it, so a session recorded in a noisy kitchen sounds like a session recorded in a noisy kitchen. Personas can only speak from the subjects you covered, so the first hour is thin and stays thin until you come back. Two gaps, named rather than buried: today's download is your memories and conversations, not your audio files, and removing the cloned voice held at our voice provider is not yet a button in the app. Personas answer from what the person actually said and speak from what is present rather than invent something to fill a silence. The limit no engineering closes: what comes out is a version of you, never you.

Timeless AI™ is the same product with no death framing anywhere near the page.

Do not buy anything to test this. Use the voice recorder on your phone. Set a timer for twenty minutes, press record, and answer three questions out loud, in a normal voice, with nobody listening.

1. What is the best day you can genuinely remember? Not the most important. The best one. The weather, who was there, what you ate.

2. What was the worst job you ever had, and what did it teach you that you still use?

3. What did a parent say so often that you can hear the exact way they said the words? Say them in their delivery, not yours.

Do not edit. Do not restart when you stumble. Stop when the timer goes.

Then check two things. How much of those twenty minutes involved death, where the usual answer is none. And how you feel afterwards, which usually lands between amused and restless, rarely grim. If your objection is the social one, answer instead the single question one specific person would actually ask you: how did you decide to move here. The awkwardness usually turns out to be about the imagined audience rather than the recording, and an audience problem is smaller, since you choose the audience.

Unless you have recorded those stories before, that file is now the only audio in existence of your voice telling them. Keep the file.

Discomfort is not evidence that something is unhealthy, and you may stop at any point. If the room goes cold when you try the sentence at dinner, drop the subject and keep recording privately.

Talking about this in ordinary rooms is not fringe either. Death Cafe, started in September 2011 by Jon Underwood and Sue Barsky Reid, on the model developed by the Swiss sociologist Bernard Crettaz, has run 24,265 events across 97 countries. People meet, usually over tea and cake, and talk (Death Cafe, count checked 30 July 2026). Nobody there is being brave. They are just not pretending.

The work asks for your best day, your worst job and the argument that changed your mind, and those three answers are the start of a free build: 50 memories, no card at any point, and no expiry. The full method for interviewing yourself is free, and what the paid plans add is set out before you record a word.

Start building your legacy Free build, 50 memories, no card required.

• The worry that you are not important enough to leave anything behind is the audience question, and it does more work than the morbidity worry does.

• The ethics of building Personas of people who have died is the half of this category that genuinely deserves the unease.

• Recording the practical things only you know how to do is for the version of the doubt where you think your life has been ordinary.

• The step by step method for creating an AI version of yourself is the plain method, with none of the reassurance.

• What to actually talk about in the first session is for the reader who has decided and now needs a first question.

• Who controls your identity after you die is the legal side of who may run a version of you at all, and the state-by-state picture on digital replicas collects the American rules.

The claims here come from peer reviewed journal articles and systematic reviews read at source (the psychology of death reflection, life review and self-disclosure, and the clinical literature on surrogate decision making), from the published text of Regulation (EU) 2024/1689 as hosted by the European Commission, and from figures published by the organisations named in them. What is not settled: Article 50 has been amended by the Digital Omnibus on AI and the published wording lags behind, the surrogate decision making evidence points two ways depending on whether the planning was structured or casual, and the mind perception research was never conducted on grief or on recorded interviews, so it is applied here as a mechanism rather than as proof. Nothing on this page is legal or medical advice, and none of this work is a will, a power of attorney or an advance directive.

Found an error on this page? Tell us through the Afterlife.ai™ contact page and we will correct the page or say why we have not.

1. Cozzolino, P. J., Staples, A. D., Meyers, L. S., & Samboceti, J. (2004). "Greed, Death, and Values: From Terror Management to Transcendence Management Theory." Personality and Social Psychology Bulletin, 30(3), 278-292. doi:10.1177/0146167203260716

2. Tamir, D. I., & Mitchell, J. P. (2012). "Disclosing information about the self is intrinsically rewarding." Proceedings of the National Academy of Sciences, 109(21), 8038-8043. doi:10.1073/pnas.1202129109

3. Kardas, M., Kumar, A., & Epley, N. (2022). "Overly shallow?: Miscalibrated expectations create a barrier to deeper conversation." Journal of Personality and Social Psychology, 122(3), 367-398. doi:10.1037/pspa0000281

4. Pinquart, M., & Forstmeier, S. (2012). "Effects of reminiscence interventions on psychosocial outcomes: a meta-analysis." Aging & Mental Health, 16(5), 541-558. doi:10.1080/13607863.2011.651434

5. Butler, R. N. (1963). "The Life Review: An Interpretation of Reminiscence in the Aged." Psychiatry, 26(1), 65-76. doi:10.1080/00332747.1963.11023339

6. Shalowitz, D. I., Garrett-Mayer, E., & Wendler, D. (2006). "The accuracy of surrogate decision makers: a systematic review." Archives of Internal Medicine, 166(5), 493-497. doi:10.1001/archinte.166.5.493

7. Wendler, D., & Rid, A. (2011). "Systematic review: the effect on surrogates of making treatment decisions for others." Annals of Internal Medicine, 154(5), 336-346. doi:10.7326/0003-4819-154-5-201103010-00008

8. Detering, K. M., Hancock, A. D., Reade, M. C., & Silvester, W. (2010). "The impact of advance care planning on end of life care in elderly patients: randomised controlled trial." BMJ, 340, c1345. doi:10.1136/bmj.c1345

9. Gray, K., & Wegner, D. M. (2012). "Feeling robots and human zombies: mind perception and the uncanny valley." Cognition, 125(1), 125-130. DOI 10.1016/j.cognition.2012.06.007. pubmed.ncbi.nlm.nih.gov/22784682

10. Hollanek, T., & Nowaczyk-Basinska, K. (2024). "Griefbots, Deadbots, Postmortem Avatars: on Responsible Applications of Generative AI in the Digital Afterlife Industry." Philosophy & Technology, 37(2), article 63, published 9 May 2024. doi:10.1007/s13347-024-00744-w

11. Prentice, D. A., & Miller, D. T. (1993). "Pluralistic ignorance and alcohol use on campus: some consequences of misperceiving the social norm." Journal of Personality and Social Psychology, 64(2), 243-256. DOI 10.1037/0022-3514.64.2.243

12. Fischer, C. S. (1994). America Calling: A Social History of the Telephone to 1940. University of California Press. The quoted line about women's use of the telephone is from the publisher's description of the book, not from the text. University of California Press listing for America Calling

13. Fischer, C. S. (1988). "'Touch Someone': The Telephone Industry Discovers Sociability." Technology and Culture, 29(1), 32-61. muse.jhu.edu/pub/1/article/888496

14. Wang, C., Chen, S., Wu, Y., and colleagues (2023). "Neural Codec Language Models are Zero-Shot Text to Speech Synthesizers" (VALL-E), arXiv:2301.02111, posted 5 January 2023. arxiv.org/abs/2301.02111

15. Regulation (EU) 2024/1689 (EU AI Act), Article 50(4), European Commission AI Act Service Desk, text checked 30 July 2026. ai-act-service-desk.ec.europa.eu

16. Regulation (EU) 2024/1689 (EU AI Act), Article 113, entry into force and application dates. artificialintelligenceact.eu/article/113

17. Death Cafe. "What is Death Cafe?" Event and country totals checked 30 July 2026. deathcafe.com/what

18. The Conversation Project. National Survey, 2018. Figures as published by the organisation, checked 30 July 2026. theconversationproject.org/about

Last reviewed: 30 July 2026. On a 30 day recheck cycle because Article 50 is affected by amendments in the Digital Omnibus on AI Regulation proposal. The page also cites survey and organisational figures that change over time.