AI can preserve a loved one’s voice and personality after death. It can also turn mourning into a subscription.
When Diego Felix dos Santos uploaded an audio message his late father had sent from a hospital bed to the voice generation platform ElevenLabs, his intention was personal. According to reporting by Reuters, he paid a $22 monthly subscription to synthesize new messages in his father’s voice, later reflecting on the output: “It feels like, almost, he’s here.”
The software mapped recorded acoustic pitch and pacing to generate new speech modeled on his father. The words themselves had never been spoken by his father in life.
Photographs, letters and recordings preserve traces of someone who has died. Generative AI can use those traces to produce an interaction. A memorial that answers back requires computing resources and may involve recurring subscription fees.
That changes the economics of remembrance. A family once bought a photograph or kept a recording. Now it can end up paying a company to keep an artificial version of someone available.
Two Kinds of Digital Ghost
Years of text messages, voice notes, emails, and family videos can provide material for reproducing aspects of a person’s cadence and conversational style.
That material is being used in two broad ways.
In one model, a living person chooses to author their own digital legacy. Platforms like StoryFile were built around video sessions where individuals record answers to specific questions about their lives. When relatives interact with the system later, the software retrieves and plays the authentic statements the person actually recorded. The user controls the archive and decides who receives access.
The second model reconstructs the person after death. Surviving relatives gather old chat logs, voicemails, and video clips, submitting the archive to a software vendor to assemble an interactive persona. In that model, the subject had no opportunity to review the material, establish permissions, or approve how the persona would speak.
That distinction governs who controls the resulting product. Someone who records their own answers has exercised some control over the archive. Someone reconstructed from private messages never made that choice, leaving decisions about how that replica is stored, managed and presented entirely to others.
The Comfort of the Simulation
For some families, the initial interaction can provide a sense of relief.
In South Korea, commercial video production services create synthetic video messages of deceased parents and grandparents for traditional ancestral remembrance ceremonies. Seoul-based startup Vaice told the Associated Press it was serving roughly 300 customers a month, based on its own reported figures, charging about 600,000 won, or roughly $390, for a short memorial video. Relatives described the clips as comforting, giving younger children a way to see and hear ancestors they were too young to remember.
That emotional response helps explain why the technology can become a consumer market. A memorial does not have to reproduce the past perfectly for someone to find it worth paying for. The commercial question begins when that experience is no longer a one-time memorial but a service that can remain active for months or years.
Bereavement does not require severing every emotional connection with someone who has died. People maintain those relationships through memory, ritual, and ongoing conversation. A conversational replica introduces a new element because the software speaks back.
A June 2026 systematic review of 30 studies on digital grief technologies published in ScienceDirect found potential benefits, including emotional solace and greater accessibility. The same review identified risks such as emotional overreliance and misrepresentation of the deceased person’s actual character. The authors emphasized, however, that empirical evidence on long-term clinical outcomes remains in its infancy, with most available research evaluating short-term user feedback.
When the Machine Starts Inventing
Conversational systems face a different challenge when answering unprompted questions.
A generative model can produce an answer that was never recorded by the deceased. If a user asks whether a deceased parent was proud of them, the system may generate an answer that sounds plausible without being something the parent ever said. Asked about an old family disagreement, it may produce a plausible narrative that was never part of the historical record.
A 2026 University of Colorado Boulder exploratory study investigating what its authors call “generative ghosts” examined user reactions among 16 participants who had lost a close friend or relative. Researchers Jack Manning and Jed Brubaker observed that factual accuracy mattered less to participants than whether the chatbot met their emotional expectations. Some participants found comfort even when the system invented details, while others found the ungrounded responses unsettling. One participant expressed concern about a relative gaining access to the tool, stating, “I just fear he would get addicted.”
That finding complicates the idea that accuracy is the product’s main measure of quality. For some users, the value may lie less in whether the system reproduces the past exactly than in whether the interaction feels recognizably like the relationship they remember. That is commercially significant because a company does not necessarily have to preserve a perfect historical record to sell a convincing experience.
Human memory is reconstructive. Repeated conversations with an ungrounded system could introduce new experiences into a survivor’s relationship with the memory of the deceased, although the long-term effect remains unknown.
The Mechanics of the Subscription Afterlife
A conventional gravestone or family photo album requires no monthly maintenance fee. Interactive digital personas do, and pricing across the sector spans a wide range.
Eternos charged $15,000 as a one-time fee to build a comprehensive interactive replica of a client’s voice, life story, and personality, drawn from 25 hours of recorded conversation before death. That was the arrangement for the company’s first client, startup entrepreneur Michael Bommer, whose legal rights to the AI version of himself were treated as a transferable asset. Separately, Eternos offered ongoing legacy account subscriptions starting at $25 per month to keep a profile accessible after death. HereAfter AI ran on smaller consumer subscriptions, historically priced between $3.99 and $7.99 per month.
The pricing reflects how differently companies have built these services, from simple life-story archives to more elaborate interactive replicas that require ongoing hosting and maintenance.
Once the replica lives on a company’s infrastructure, remembrance acquires the basic mechanics of a software product. There is an account to maintain, data to store, computing to pay for and access to control.
This structure introduces a fundamental tension over ownership. A family can possess the original recordings or messages while a company controls the software through which those memories become interactive. In practice, the recordings, generated model, account and contractual rights may be governed by different rules. For a subscriber with no independent local copy of the recordings, canceling the service can leave the family dependent on the provider’s retention and access policies. Where the service uses recurring billing, the company continues earning revenue while the customer maintains access.
The Problem of Company Mortality
A digital persona’s continued availability depends on the infrastructure and organization supporting it. Startups in this sector face the same commercial pressures as other technology companies, but their products raise a different problem because customers may expect them to remain available for decades.
StoryFile filed for Chapter 11 bankruptcy protection in May 2024. As of an April 2026 court filing, the company remained in Chapter 11, with its interim CEO stating in a sworn affidavit that he believed StoryFile had sufficient business prospects to fund a reorganization plan.
HereAfter AI’s own website indicated in 2026 that its consumer service had ended, directing users to a support email to request their recordings. Eternos rebranded in late 2025 as Uare.ai, shifting its focus toward AI tools for living professionals.
When a digital legacy company closes or restructures, practical problems follow for families: who holds the rights to the underlying data, whether voice models can be exported, and what happens to private correspondence if company assets are liquidated.
Data Ownership and the Law
A deceased person’s archive rarely contains only their own data. A decade of message history holds the names, medical details, arguments, and photographs of living children, siblings, and friends. A family uploading years of private messages may be handing a company information about people who never agreed to have it used for an AI recreation. The legal treatment depends on the jurisdiction, the provider and the information involved.
In the United States, statutory protection is fragmented by state and generally addresses commercial uses of a deceased person’s identity rather than every private use of a digital replica. In December 2025, New York enacted legislation under Senate Bill S8391 expanding post-mortem protections. The statute creates distinct rules for deceased personalities whose likeness had commercial value at death, and establishes specific consent requirements for using digital replicas of deceased performers in audiovisual works and sound recordings.
California takes a different approach. Assembly Bill 1836, effective January 1, 2025, amended Civil Code Section 3344.1 to explicitly include digital replicas within California’s 70-year post-mortem right of publicity for deceased personalities in commercial works. The statute also contains exemptions covering specified uses involving news, public affairs, political activity and expressive works, leaving some non-commercial representations unsettled.
Neither statute directly answers every question raised by a private family griefbot. Their protections are tied to particular jurisdictions, people and uses, leaving many ordinary cases unresolved.
The Battle Over Posthumous Consent
Following the January 2026 death of cartoonist Scott Adams in Pleasanton, California, an artificial intelligence video avatar surfaced online, broadcasting commentary in Adams’ likeness and voice. As reported by Business Insider, the avatar’s creator, venture capitalist John Arrow, defended the project by citing Adams’ repeated public statements giving permission for a posthumous AI: “I’ve looked and looked and can find no evidence of any type of revocation. If there was anything that suggested this is what he didn’t want, I would stop.”
The Adams family and estate rejected that justification, demanding that Arrow stop using Adams’ likeness and calling the digital twin an unauthorized fabrication. Karen North, a USC professor who studies digital social media and psychology, pushed back on Arrow’s defense, noting that conversational remarks, even sincere ones, “may not be legally binding in ways contracts and intellectual property rights are legally binding.” North also observed that calling these systems avatars softens what they actually are, allowing realistic imitations to surface while a family is grieving, potentially expressing views the deceased person never held.
Whether the Adams estate has a claim under California’s Section 3344.1 remains unresolved. Legal scholar Betsy Rosenblatt told Business Insider that a digital replica not selling anything becomes “more likely to be considered a First Amendment protected expression” for its creator, calling that defense “not a slam dunk, but a stronger argument.” That analysis was offered for an anonymous creator, whereas Arrow is publicly named. The AI Adams account identifies itself as artificial intelligence and does not appear to solicit money. Those facts could support an argument that the project falls closer to the expressive side of the statutory exemptions. The dispute has not been tested in court.
The Adams dispute tests the entire ownership hierarchy: if an individual expresses support for digital preservation while living, but their family objects after death, who has the final authority to decide what the deceased becomes?
What Remains Uncertain
The Adams dispute highlights a problem that many families could face without public attention. Adams left behind years of recorded commentary regarding his wishes. Few individuals leave an equivalent record, leaving families and companies with no documented instructions.
That leaves consumers unable to evaluate whether the service they are purchasing is durable enough to safeguard a person’s digital identity. The market itself is difficult to measure, as broader commercial forecasts for artificial intelligence avatars combine customer-service bots, gaming characters, and virtual influencers.
The clinical evidence is still catching up with the technology. The CU Boulder study represents an early examination of user experience, and follow-up research into long-term mental health outcomes is only beginning.
Research involving children is even more limited. Little longitudinal evidence exists on how children process interactions with simulated personas of deceased parents over years of development.
The Reality of the Product
An interactive digital persona depends on infrastructure and a service that continue to operate.
For Lee Geon Hui in Seoul, that meant watching his father moved to tears by a short video of a grandfather who died decades earlier. Diego Felix dos Santos continues paying to interact with a synthetic voice modeled on an audio message sent from his father’s hospital bed. In California, the family of Scott Adams is publicly contesting software that broadcasts political commentary in a dead man’s voice without their consent.
These experiences reflect different needs, but they share the same underlying structure. A family no longer simply possesses a memory. It relies on an external platform to generate an interaction, leaving the continued existence of that voice dependent on servers, commercial accounts, and companies that may not outlive the people they recreate.
