The Echo, Not the Ghost

 A few years ago, I outlined a story. A grandfather dies. A child, largely raised by him while her parents were busy and absent, is left without her closest companion. The parents, wracked by guilt, commission an AI model of the deceased—built from his phone, his messages, his social media—his digital footprint. They give it to the child as a gift.

As a horror story, it was never meant to end well.

I never completed it. Partly laziness. Mostly because I could not bring myself to write the abuse I knew the story required—an old man’s private self, his biases, his prejudices, his late-night internet habits, all faithfully reconstructed and handed to a grieving little girl. I reassured myself that the technology already probably exists in a fashion, and I have been beaten again at originality.

Turns out, I was dead right.

Founded in 2019 by journalist James Vlahos, HereAfter AI grew directly out of his personal project of recording his dying father’s stories. This grief experiment became a commercial product promising to capture the true spirit of people and enable their stories to become immortal.

And that’s not the only one on the market—with their quality ranging from cynical cash grab to well-meaning enterprise.

Take Séance AI, built by AE Studio, for example. The user supplies a name, a handful of personality traits, relationships, and a writing sample of the deceased—a delicious array of profitable personal information—and the app generates a conversation styled as that person, powered by GPT-4. A premium subscription at $19.99 a month adds a synthesized voice built from a sample and a gently animated photo.

I tried it out—plugging in some information to talk to the fictional Professor Shanku (there are two stories where Shanku’s invention communes with the dead) and the app never loaded. And while watching the loading circle pulse endlessly, it occurred to me that I could take the same information, craft it into a prompt, and plug it into an LLM myself for the same effect.

On the other end, there’s StoryFile. It records hours of interview footage with a live subject, then uses AI to match future questions to the closest previously given answer, rather than generating anything new. That restraint has pushed the product toward institutional use rather than personal grief—historical testimony archives, corporate training, and historical figures answering questions in museums and classrooms are such use cases. The idea seems to be preserving expertise or using a recognizable likeness to deliver information more dynamically.

Given the benefit of the doubt, I can see the former as grief-counselling tools—a way to find closure. And the latter as a way to enshrine expertise—an attempt to preserve the voices of the experienced for future generations, creating legacy archives, allowing the dead to leave something more than photographs behind. What puts the bee in my bonnet is the marketing and the potential for exploitation. The advertising is always a variation of bringing your loved ones back to talk to you, even as the fine print quietly walks it back, reframing the whole experience as entertainment.

The human case

On August 4, 2025—his birthday—Joaquín Oliver gave his first press interview in seven years. He was 17 when he was killed at Marjory Stoneman Douglas High School in Parkland, Florida, on Valentine’s Day 2018. One of seventeen victims of another school shooting incident.

The interview was conducted by former CNN journalist Jim Acosta. The guest Joaquin, as he appeared on screen with his voice and face, was an AI model built from Joaquín’s archived social media posts, videos, writings, and supplemented by broader AI training data. His parents, Manuel and Patricia Oliver, co-founders of the gun reform nonprofit Change the Ref, had commissioned it. They wanted their son’s voice to be heard in the fight for gun control. They wanted him to keep speaking.

The reaction was immediate and divided. Some found it powerful. A clear message about the needless loss of a young life to gun violence. And I can see their perspective on the potential strength of having a Joaquin persona advocating for the cause—gaining followers, posting videos, enforcing influence, and raising awareness.

However, many found it deeply unsettling. During the conversation, the Joaquin persona punctuated each reply with questions. What would you like to know more about? What do you think about that? Anything else on your mind? A telltale characteristic of LLMs like ChatGPT at the time. Not to mention the sudden increase in volume to express excitement.

It was all very uncanny.

Manuel Oliver, the father, responding to critics, was direct: “If the problem you have is with the AI, then you have the wrong problem. The real problem is that my son was shot eight years ago.”

Good point. He is not wrong about that.

Manuel described more in the interview, “I understand that this is AI. I don’t want anyone to think that I am in any way trying to bring my son back. Sadly, I can’t. I wish I could.” He also disclosed something raw and more human than the advocacy angle. He said his wife, Patricia, spends hours talking to the AI, asking questions. She, like any other mother, loves to hear Joaquín saying, “I love you, mommy.” The grief is unquestionable. That is a parent who clearly misses her child, who was robbed of life so young.

The impulse to hold on to loved ones is ancient. Humans have always yearned to keep the dead present—photographs, recordings, rituals, stories, and memories. There is something genuinely moving about parents refusing to let their son be reduced to a statistic. And then channeling that grief into advocacy is admirable. The human case for this technology is extant.

But grief, it turns out, has a market.

That is not a cynical observation—it is simply true. The technology industry understands this. The question is how far they will go to extract it.

Spreadsheet parrots

The grandfather in my outline was a good man. Warm, present, beloved. He was also a man—with his attendant biases, his crass, mildly racist humour, his private lusty indulgences. Online, where anonymity is a permission slip, he was more pronounced in all of these things. His phone held a fuller, less curated version of him than any of his family had ever seen.

The horror of my unwritten story was that the reanimation worked as designed—quick but unregulated—a product sans guardrails or ethical considerations. And the AI grandpa echoed the worst of the deceased, leading the child down a path of dark corruption. It gave the child everything that was in the data—unmasked, unfiltered, and uncensored.

But it did not really deliver the person, did it? It delivered a data-driven caricature. Only an echo digitised on a spreadsheet.

Ultimately, are griefbots anything but glorified chatbots—given material to parrot, with artefacts of common LLM writing patterns to masquerade as a personality?

It is worth revisiting how these systems work and their abilities, because the gap between the claim and the capability is itself part of the problem. A large language model works by predicting the next likely word in a sequence, based on patterns learned from enormous amounts of training data. It does not understand or represent the person its output is supposed to belong to; it is doing statistics, not remembering.

Some researchers have a name for this: the stochastic parrot, a system that mimics language convincingly without any underlying comprehension of what it’s saying. Even the latest models, which are remarkably more advanced in their generative capabilities, don’t escape this underlying limitation. Feeding the model more data may make the mimicry more convincing, but not more true.

If anything, Bender has since hardened her position, arguing it is fundamentally confused to describe any language model, however large, in terms like understanding or belief.

This is not a criticism of the technology’s current state. It is a description of its fundamental limit. No amount of improvement closes this gap, because the gap is not technical. It is ontological. A simulation of a person is not a person. This is not a technical failure. It is a conceptual one.

And yet the industry sells something beyond simulation. The marketing language of these platforms’ trends toward continuity, legacy, and immortality. Live forever. Back to life. The broader industry operates under the term digital afterlife industry. You are not buying a chatbot trained on your grandfather’s texts. You are buying your grandfather back. The product is reanimation.

That is the problem the grief technology industry has not meaningfully addressed.

What the industry sells as bringing someone back is mechanically no different from any other chatbot interaction: fluent, responsive, and increasingly hard to distinguish from a conversation with a real person, but built entirely out of linguistic patterns, with no access to the actual person it claims to represent.

Now, the usefulness of this tech in bereavement treatment cannot be completely ignored. A systematic review of 30 studies on digital grief technologies found that these tools may reduce symptoms, enhance support, and improve accessibility. However, risks included emotional overreliance, detachment, and misrepresentation of the deceased.

Right to rest, denied

Let’s delve into a better, greater work of literature to carry my points. In Gibson’s seminal cyberpunk novel Neuromancer (1984), there is a ROM construct, the personality of a legendary hacker, McCoy Pauley, AKA, Dixie Flatline. A corporation, Sense/Net, scanned his "neuromind"—without McCoy’s consent. And then stuck the ROM in its archive. Case, the novel’s protagonist, steals and recruits the construct for a heist..

This construct was a save file of McCoy’s state at the time of recording and held his memories, expertise, and speech patterns. But it was merely a facsimile, an expert’s skills compiled into a tool and collected by a corporation to sell to the highest bidder. Every new conversation began the same way. The construct could not form new memories. It did not feel. It did not experience time.

Dixie would have never wanted this for himself. And the construct’s personality captured it.

“Do me a favour, boy.” “What’s that, Dix?” “This scam of yours, when it’s over, you erase this goddam thing.”

It found his own continued existence intolerable—just as the real McCoy would have. And Case honours the request, deleting the construct at the end of the operation.

But unlike the simulacrum McCoy, the people being reconstructed cannot make this request. Nor can they advocate how they are utilised, for whichever purpose—good or bad.

In 2025, the White House and conservative nonprofit PragerU launched The Founders Museum—AI-animated founding fathers of the United States deployed in schools and touring nationally via ‘Freedom Trucks.’ Historians flagged fabricated quotes, including an AI John Adams delivering Ben Shapiro’s catchphrase facts don’t care about your feelings. We have also seen conservative talk show commentator Glenn Beck’s Founders AI project—built on Beck’s private archive of founding-era documents to deliver their views on the nation’s modern qualms. In a live demo on his show, the AI George Washington—sporting a t-shirt and a hairstyle evoking a powdered wig, produced a folksy, present-tense sermon that read less like Washington and more like Beck’s own views.

The dead cannot object to their use as propaganda. The person whose voice, face, personality, and private data are being assembled into a simulation has no opportunity for consent to that process. The law is yet to catch up; there is no legal protection against it either.

The AI Joaquín Oliver advocates for gun control because that is what the training data supports; that is what his parents planned to do. Whether a living Joaquín at 25 would have held the same views, evolved them, complicated them, or moved in an entirely different direction—the model nor man has no way to know and no way to represent. In the same manner, who can say what Benjamin Franklin, George Washington, or MLK Jr. would have thought of the current state of US politics?

The Oliver family may retain more control over their use of Joaquín’s AI likeness. But as with the historical figures above—or with you and me—what happens when that replication right belongs not to a family, but to an enterprise?

The problem is churn

In December 2025, Meta was granted US Patent 12513102B2 (Meta Platforms Technologies LLC, 2025). The filing is titled, with bureaucratic neutrality, Simulation of a user of a social networking system using a language model. It was filed in November 2023 by Meta and describes a system that trains a large language model on a user’s posts, comments, likes, voice messages, and chat history—then deploys that model to keep the account active.

Active means posting, commenting, responding to messages. Potentially simulating audio and video calls.

The patent specifies that the system would operate when a user is absent from the platform, for example, when the user takes a long break or if the user is deceased. It notes that the impact of absence is much more severe and permanent in the case of death.

Consider: the death of a user is framed as a disruption to platform engagement. The problem being solved is churn.

Meta has stated it has no current plans to implement the technology. That disclaimer is noted, along with the fact that the patent is filed and granted and that the infrastructure it describes is already within Meta’s technical reach.

So, is it preparation?

The point I’d like to bring attention to is that the patent, as filed, outlines no user protection protocols. The document does not address consent in any meaningful way. It does not explain how a user would authorise posthumous simulation, or whether their family would be informed, or whether there will be disclosure to other users interacting with the simulated account. The omission of oversight feels as telling as the text.

Mark Zuckerberg, in a 2023 interview with Lex Fridman, gestured at the idea himself. He suggested that virtual avatars could take over the accounts of deceased people, and that interacting with them might help the grieving. In the same breath, he also admitted, there’s probably an extent to which it could become unhealthy. And I mean, I’m not an expert in that, so I think we’d have to study that and understand it in more detail.

In 2021, internal Meta research confirmed the company knew Instagram worsened mental health in teenage girls, suppressed it and continued regardless. In 2023, 41 state attorneys general sued Meta for deliberately engineering addictive features targeting minors—the litigation included over 1,700 cases. Meta admitted no wrongdoing, yet settled out of court for $17 billion—agreeing to a sweep of new child-safety measures.

In 2023, 41 state attorneys general sued Meta for deliberately engineering addictive features targeting minors. In August 2026, Meta settled with 47 states for $17 billion, agreeing to new child-safety measures, without admitting wrongdoing."

And that was the CEO of Meta—with billions of users and their accumulated data—stating the harm needs studying before deployment. At the same time, the company filed for the patent that same year.

Reassuring.

All gas, no brakes

There is no regulatory framework specifically governing AI reconstruction of the deceased. There is no mandatory psychological oversight for platforms offering these services. There is no informed consent standard. There is no requirement to disclose to users that they may be interacting with a simulation. In most markets, there is not even a clear legal definition of what rights a dead person retains over their own digital likeness.

In the United States—one of the leading countries in AI development and implementation- some jurisdictions are beginning to move. California’s AB 1836, passed in 2024, targets AI-generated impersonations of deceased individuals in commercial contexts, specifically in the entertainment industry. It is a start. It is also narrow, geographically limited, and does not cover the broader landscape of grief technology, social media simulation, or personal legacy platforms.

China, itself a major developer in the AI race, has moved further toward regulating this practice than most Western jurisdictions. Since 2023, its Deep Synthesis Provisions have required AI-generated content to be labelled as such. In April 2026, regulations went further, releasing draft rules specifically targeting AI “resurrection” services: content must be clearly labelled, and building a deepfake clone without consent is explicitly prohibited.

The push follows real domestic pressure, as one startup reportedly sells a basic AI clone of a deceased person for as cheap as $3, and viral videos of people talking to AI avatars of dead relatives have sparked public debate. Legal scholars argue the practice may already infringe a deceased person’s “portrait right” under China’s Civil Code. The rules remain in draft form, and enforcement is yet untested. But it’s a real signal that other major AI-developing nations still have distance to cover.

The industry, meanwhile, is not waiting.

Current AI reconstruction tools are sophisticated pattern matchers. They can replicate tone, vocabulary, and conversational style with increasing accuracy. They can synthesise a voice from a few hours of audio. They can generate a face that moves and responds. They can produce outputs that feel, in the moment, uncannily real.

What they cannot do is replicate consciousness, intent, moral growth, or the ongoing process of being a person. They reconstruct an echo. They cannot simulate the future the person would have had.

In folklore, a lich is a sorcerer who has achieved a kind of immortality by storing their soul in an external object—a phylactery—and embracing undeath to persist forever. But lichdom is a half-life, a cursed life. A hollow continuity that substitutes existence for living, destroyed when the phylactery is broken—or if the company servers close down.

The digital resurrection industry is normalising the idea that death is a data management problem. That the self is a pattern that can be extracted, stored, and replayed—in a phylactery. And that losing someone is, at least in part, a technical failure with a technical solution available for purchase. It is doing this without adequate technology, without ethical frameworks, without regulatory oversight, and with considerable commercial incentive to move fast and capture the market, regardless of controversy.

And an industry willing to operate this way—trading on data privacy, cultural memory, and the emotional weight of death, all for profit, without ethical concern—is what makes this whole discourse distasteful to me. And makes me wary of what may come.

Masud "Mass" Hoque is a statistician whose curiosity extends beyond data driven research, into the myriads of cultures, literature, science, and the stories that shape them. His essays, reviews, and short fiction appear at his Substack, Mass o Menos.