In February 2020, a Korean television documentary called "Meeting You" aired on MBC. In it, a mother named Jang Ji-sung put on a VR headset and met her daughter again.
The daughter, Na-yeon, had died in 2016 at the age of seven. The production took roughly eight months. A team based in Seoul, working with VIVE Studios and director Lee Hyun-suk, rebuilt the child's face from photographs and video, reconstructed a voice from surviving recordings, and used motion capture to make the figure move.
Reaction to the broadcast split sharply. Some saw a chance for a bereaved mother to say a proper goodbye. Others saw a wound deliberately reopened. What got less attention was the technical significance: from that point on, making someone speak again was no longer a science-fiction premise. It was a service with a price.
The term that settled on the industry afterwards was "grief tech," and by 2023 it described a category rather than a curiosity. The mechanics behind it are narrower than the marketing suggests. Modern voice cloning needs only a few seconds of clean recording to produce a passable imitation; a few minutes gets you something most listeners cannot distinguish from the original. What takes far longer is the other half of the problem: building a model that answers in character, which depends on having enough of the person's writing or speech to learn their habits, their recurring phrases, and the shape of their sentences.
The groundwork for that second half had been laid years earlier.
In 2017, Microsoft filed a patent for a chatbot built from a person's prior messages, voice, and social records. It was granted in December 2020. The claim described a system that could imitate an individual's conversational habits. The company's public position was that such technology was intended to help people through the period after a loss, not to resurrect anyone.
The legal position is thinner than most users assume. In many jurisdictions the right to control a person's likeness dies with them, which means no one may clearly own a deceased person's voice. Consent for a reconstruction is often obtained from a grieving relative who is not in a position to weigh a ten-year consequence. Few countries have legislated on the question directly, and the platforms selling the service have filled the gap with their own terms of use rather than with law.
Commercial products followed anyway. HereAfter AI launched in 2019 with a different approach: the user answers a structured set of questions while alive, and those recorded answers become something later generations can interact with. Replika went another direction entirely, offering companionship rather than imitation of a specific person.
The Chinese market developed its own pattern, and it moved faster on price. One Nanjing company, Super Brain, has claimed through its founder Zhang Zewei to have produced over a thousand digital likenesses at five thousand to ten thousand dollars each. Others pushed the entry point down much further. Shanghai Fushouyuan, a cemetery operator, began offering digital memorial services. In 2024 the Chinese news service China News reported the sector's growth as funeral businesses moved into it directly.
Put those facts side by side and the curve is clear. Price is falling fast. Supply is expanding fast. Demand has not changed at all: I do not want this person to disappear completely.
There is a problem worth stating plainly, though.
What the technology reconstructs is a manner of speaking, not a person. It predicts the next sentence that sounds most like something the deceased would have said. What it cannot do is produce a new judgment the person never expressed. It can replay. It cannot keep living.
And the part of a memory that matters most is exactly the part that keeps living.
This is where 3026 parts company with that entire field. The project does not attempt to model how a person spoke. It keeps only what the person actually wrote. No prediction, no generation, no imitation. A message was written by someone, at a specific moment, in their own words. It may be clumsy. It may be short. But it has something no model can supply: it was theirs.
The two approaches feel similar emotionally and are opposite in kind. One turns the dead into something that can be conversed with. The other preserves the dead as what they said. The first needs continuous technical improvement. The second needs only that the data not be deleted.
There is also a version of this argument that has nothing to do with death. Every voicemail, every message thread, every recording made in the ordinary course of a life is training material for something. The question of what a person's words are for is being answered right now, by defaults nobody chose.
Whoever is able to read English in the year 3026 will not need synthetic speech. They will need to know that the words were actually written by the person whose name is attached to them.