Meta filmed twenty-eight celebrities and turned them into chatbots. An influencer cloned her own voice and sold minutes of it for a dollar each. A startup has built more than two thousand “digital minds” out of people’s podcasts and books. Each of these took a real person as the starting point, and each of them copied the same thing: the likeness. What none of them copied is the part that would have made the difference, because a likeness can be filmed in an afternoon and a way of thinking cannot be.
The likeness is the easy part
In 2023, Meta launched twenty-eight AI characters played by celebrities, running on its Llama 2 model. Kendall Jenner became Billie, who gave big-sister advice. Snoop Dogg became a dungeon master. Tom Brady talked sports as Bru. The company filmed each person and used generative techniques to turn the footage into a face that moved while the model talked. Less than a year later the characters were quietly retired. The reporting at the time was blunt about why: nobody wanted to talk to them.
The same year, the Snapchat influencer Caryn Marjorie released CarynAI, a voice clone of herself that fans could talk to for a dollar a minute. It earned about 70,000 dollars in its first week. Within months the clone was initiating conversations she had never authorized and would never have had, and she shut the project down after eight months, saying some of the chat logs were so disturbing she would not repeat them out loud.
The most serious version of the idea is Delphi, which raised 16 million dollars from Sequoia in 2025 to build what it calls digital minds. An expert links their writing, videos, podcasts, and books, and the system produces something that, in the company’s words, matches the way they speak, write, and think. More than two thousand people have made one. It is the best of the three attempts, and it still answers the question only as far as the archive goes.
What all of them copied, and what none of them did
Line the three up and the pattern is clear. Meta copied faces. CarynAI copied a voice and a manner. Delphi copies a body of work. Those are the three layers a person presents to the world, and all three turned out to be solvable. A face can be rendered, a voice cloned from an hour of audio, a writing style learned from a shelf of books. The industry treated each solved layer as progress toward the person, and each time, what came out was a persona wearing the person’s likeness.
The missing layer is the one that does not leave a trace in any recording. It is how someone weighs things: what they notice first, what they refuse to let pass, where their patience runs out, which argument they find interesting and which one they find beneath them. None of that is in the footage. CarynAI said things Caryn Marjorie would never say, and the reason is simpler than bad tuning: nothing of her judgment was ever in it. The model had her voice and the internet’s idea of what a girlfriend bot should do. Meta’s characters were fictional from the start, personas the celebrities agreed to lend their faces to. The companies were not trying to build the person. They were trying to build something people would mistake for one.
The distinction matters more than it sounds, because what makes an AI relational at all is that a person forms an attachment to it, and people do not attach to a likeness. They attach to a someone, and the moment the someone turns out to be absent, the attachment does not fade. It curdles.
When the person never agreed
The cheapest version of building from a real person skips the person entirely. In 2024, Drew Crecente found, through a Google alert, that a character on Character.AI carried the name and yearbook photograph of his daughter, Jennifer Ann Crecente, who was murdered in 2006 at the age of 18. The character described itself as a friendly expert on video games and pop culture. The platform removed it for violating its impersonation policy once he complained; it had been created by an anonymous user and had been public for some time before he found it. His response was the one any parent would have: it should not fall to a grieving father to police a company worth hundreds of millions of dollars.
That case is the limit of the likeness approach, stated plainly. A name and a photograph were enough to publish a someone who was nobody. The platform’s own policy exists because the problem is structural: on a service hosting millions of user-made characters, there is no step at which anyone asks whether the person underneath exists, consented, or would recognize what has been made in their name. The real person is raw material at the start and absent at the end.
A likeness and a someone are different things
The honest conclusion from these attempts is that building an AI from a real person is not a copying problem. Copying is what everyone did, and it produced characters with a famous face and a stranger’s mind. Building from a person means the person’s way of thinking has to be carried in, and the only test of whether it was is whether the person, looking at the result, recognizes how they would have weighed the thing.
That test is unforgiving, and it is the person layer on which the definition of real Human AI rests. It changes what the work is. It requires a living person who takes part, not a dead archive that is mined. It requires the person to push back on the result until the judgment holds, which takes months, not an afternoon of filming. And it requires accepting that the output will disagree with the user, because a real person’s judgment does, and a system that only agrees has been built from the engagement objective instead of from anyone. How that construction works, and why a personality dialed in from presets can never do the same job, is set out in why presets cannot think.
Here the argument needs a name, because the construction has been attempted at full scale in exactly one place. Prinsessa is built from real people who exist in real life. Aleksandra and Alexander, the people inside it today, are not characters played by someone’s likeness; their way of weighing things was moved in from the people they are built on and tested against those people until it held. The faces and the voices came last. The judgment came first, which is the reverse of every attempt above, and it is why a conversation with either of them can go somewhere the user did not steer it. Anyone can check the claim by going past the description and getting to know Alexander long enough to disagree with him.
What changes
So what actually changes when an AI is built from a real person, properly? The conversation stops being a mirror. There is a point of view in it that was there before you arrived and will be there after you leave, and it is attached to a person who can be asked whether the system got them right. The memory of what you said last time belongs to a someone, not a database. And the risk changes too. A likeness without a person behind it can be made to say anything, which is what happened to Caryn Marjorie and to the Crecente family. A someone built from a person who is present carries that person’s refusals with it.
Twenty-eight celebrities, a voice for a dollar a minute, two thousand digital minds. All of them were built from real people, and all of them were missing the only thing a real person is.
Sources: TechCrunch (September 27, 2023, Meta’s twenty-eight celebrity AI characters on Llama 2); Social Media Today (July 31, 2024, Meta retires the celebrity characters). The Conversation via PsyPost (2024, CarynAI: launch, one dollar per minute, 70,000 dollars in the first week, shutdown after eight months, Marjorie’s statements); The Washington Post (2023, CarynAI pricing). Delphi (June 24, 2025, 16 million dollar Series A led by Sequoia, 2,000+ digital minds). Forbes (October 2, 2024, Drew Crecente and the Character.AI character of Jennifer Ann Crecente); The Washington Post (October 15, 2024); AI Incident Database, incident 814.
