Digital immortality

Digital immortality is the persistence of a person in computational form after biological death. The term covers two things that share almost nothing except a marketing vocabulary. The weak form builds a conversational model from what a person left behind — messages, recordings, video, social media — and produces something that talks like them. The strong form is Mind uploading: running an emulation derived from the structure of the person's brain, which would on some accounts be the person. The weak form exists and is sold commercially. The strong form has no working demonstration at any scale.

§The weak form

The idea entered wide circulation in 2016, when a chatbot built from a dead friend's text messages was reported at length and read by many as a glimpse of something larger.¹ A reconstruction is a stylistic model. Given enough text and audio, a language model can be conditioned or fine-tuned to reproduce a person's turns of phrase, opinions, and voice; add synthesized speech and a generated face and the output is a video call with a dead relative. Nothing about the underlying brain is involved, and the model has access only to what was recorded. Where the source data is thin, the system fills gaps by generalizing from its training corpus — which means it will confidently produce statements the person never made and could not have made.

The lineage runs back through lifelogging, the archival cousin of the self-tracking practices described under Quantified self. Gordon Bell's MyLifeBits project at Microsoft Research digitized a life's worth of documents and recordings in the 2000s on the premise that comprehensive capture would preserve something.² Martine Rothblatt's Terasem Movement, an offshoot of organized Transhumanism, promoted "mindfiles" — curated personal archives — as a substrate for later revival, and commissioned the Bina48 robot as a demonstration. Ray Kurzweil has described building a conversational replica of his late father from letters and documents, and treats it as a partial reconstruction rather than a restoration. None of this touches the biological storage of memory that a Memory prosthesis attempts to interface with; the archive is external throughout.

1. 2016
   The Roman bot
   Eugenia Kuyda trains a chatbot on text messages from her friend Roman Mazurenko after his death. The Verge reports it at length, and her company later launches the companion app Replika.
2. 2017
   Dadbot
   Journalist James Vlahos builds a conversational agent from an oral history recorded with his dying father, later founding a company offering the service commercially.
3. 2021
   Project December
   A man's use of a GPT-3-based service to simulate his deceased fiancée is reported in detail, prompting debate about consent and the emotional effects of such systems.
4. 2021
   A patent draws attention
   Reporting on a Microsoft patent covering conversational agents modelled on a specific person from their social data prompts public debate; the company says it has no plans to build the product.
5. 2024
   Ethics catches up
   Researchers publish design frameworks for 'deadbots', proposing consent requirements, dignity constraints, and exit procedures for the digital afterlife industry.

§The digital afterlife industry

Products fall into three groups. Pre-recorded interactive testimony records answers to a large question set while a person is alive and retrieves the closest match on demand; the USC Shoah Foundation's Dimensions in Testimony project, which lets visitors question Holocaust survivors, works this way and generates nothing — every answer is something the person actually said. Curated generative systems combine an interview archive with a language model, which is what most commercial services offer. Uncurated generative systems train on whatever data relatives supply, with no involvement from the person being modelled and no consent at all.

The economics matter. A reconstruction requires ongoing hosting, so the service is a subscription whose lapse ends a relationship; and because the models sit on the most intimate archive a family possesses, the sector inherits every unresolved problem in consumer data governance. Carl Öhman and Luciano Floridi have argued that the industry's structure, in which the dead become an asset class managed for profit, should be evaluated the way archaeology and funerary practice are — with duties to the deceased that are not reducible to the interests of the living.³

Not a continuation

A reconstruction has no memories, only records of memories; no experience, only generated text. Whatever one thinks of Substrate independence, nothing in these systems has the causal relationship to the original brain that even the most permissive theory of Personal identity and continuity requires. Calling them immortality is a category error, not an overstatement of degree.

§Consent, dignity, and postmortem data

The people being modelled are mostly not in a position to object. Most jurisdictions give the dead weak or no data-protection rights: the European GDPR does not apply to deceased persons, leaving the matter to member states, and in the United States protection depends on state right-of-publicity law, which was written for celebrity endorsement rather than for grief. Some recent statutes aimed at synthetic media, including Tennessee's 2024 law on voice and likeness, extend protection in ways that may reach griefbots incidentally.

Ethicists working on the problem have proposed concrete constraints rather than prohibition: obtain consent from the person during life for the specific use; require that the system disclose its nature; provide a dignified way to retire it; and prohibit its use to advertise to bereaved users.⁴ The last is not hypothetical — a system with a grieving user's attention is commercially valuable in a way that invites exactly the wrong incentives, a concern that overlaps with Mental privacy and with the broader question of who controls intimate inference.

There is also a scale argument. Modelling of social platform data has projected that profiles of deceased users will number in the billions within this century, making the accumulated digital remains of the dead a substantial and largely ungoverned archive.⁵

§What the evidence says about grief

Little, so far. Grief research has moved away from the older assumption that healthy mourning requires detachment; continuing-bonds theory holds that maintaining a relationship with the deceased is normal and often adaptive.⁶ That gives a prima facie case that interactive reconstructions could help some people. Against it sits the clinical category of prolonged grief disorder, formally added to psychiatric diagnostic manuals in the 2020s, and the concern that a system which never refuses contact could sustain avoidance of the loss.

As of 2026 there is no substantial trial evidence either way — no randomized comparison of griefbot use against standard bereavement support, and no longitudinal data on outcomes. Published work is dominated by case reports, qualitative interviews, and ethical analysis. Claims in either direction outrun what has been measured.

§The strong form

Actual digital immortality in the sense advocates intend requires Whole brain emulation: a scan of the specific brain at a resolution sufficient to reconstruct its causal organization. Nothing on that scale exists. As of 2026 the destructive methods of Connectomics have reconstructed human cortex in volumes of about a cubic millimetre, roughly a millionth of a brain, and no nervous system of any size has been emulated. Because the technology does not exist, the practical form of the strong claim is preservation — Brain preservation and Cryonics — on the argument that keeping the information available is a precondition for any later attempt.

Even granting the engineering, the strong form inherits the unresolved questions: whether an emulation would be conscious, addressed under Machine consciousness and Neural correlates of consciousness; and whether it would be the original person, addressed by the The teleportation problem. A reconstruction avoids all of these by not making the claim.

§Why the conflation matters

Marketing a reconstruction with the vocabulary of continuation misleads in a specific way: it suggests that recording more data now brings a person closer to surviving later. It does not. No quantity of text captures synaptic structure, and the two forms are not points on a single scale but different kinds of thing. The confusion also runs the other way, letting critics dismiss serious work on emulation by pointing at chatbots.

The open question in the near term is not technical but institutional. Nobody has established who owns a dead person's likeness once it is a running model, who may switch it off, what a person can bind their heirs to in advance, or whether a right to be forgotten survives death — and services are being sold while the answers are absent.

§See also

• Mind uploading
• Whole brain emulation
• Personal identity and continuity
• Brain preservation
• Cryonics
• Machine consciousness
• Mental privacy
• Substrate independence

§References