Watch the clip and your body decides before you do.
A girl sits at a desk. She smiles. She looks down, and her hand starts to move across a notebook, as if she has just found the next line. It runs twenty-three seconds. You take her in as present, here, alive in the room. Then the caption says she was born in 1929 and in a kinder world would have turned ninety-six, and you remember that she died of typhus at fifteen, in a camp, a couple of months before the war ended, and that the diary she kept in a hidden annex was published two years after she died, and later staged and filmed.
The clip leaves one thing out. There is only one real film of her, and it runs about twenty seconds. It comes from Amsterdam in July 1941, a neighbor's wedding, and in it she leans out of an upstairs window to watch the bride and groom pass. She is not at a desk. She is not writing. She does not look at us. She is a girl leaning out to see a street for a moment. That is the whole moving record of her. Everything else, the smile, the downward glance, the pen, was made by a machine.
When we worry about AI and the past, we mostly worry about fakes: the lie that wears the face of a fact. And there are lies. There is a boy on a memorial page who "liked feeding the family goat," said to have died in Treblinka, though no victim by that name appears in the names database of the murdered. There is a girl "racing the wind with flushed cheeks and wild dreams," supposedly lost at Auschwitz, and the registry has no one by her name either. Someone invented them to be mourned and shared, and a whole content economy runs on children like these. It is a real problem, and not an easy one: the fakes multiply, they travel, and each one wears down a little more trust. But a lie of that kind is at least the danger we know how to name. It asserts a false fact, and a false fact can be checked. You search the registry of the murdered. The name is not there. The post is still a lie and may still spread, but you know exactly what is wrong with it.
The clip asserts nothing false. That is what makes it worse.
The people who made it were careful. They used photographs already in the public domain. They did not make her speak; the man behind it said that making her speak "would be crossing a line." Everything the clip claims as fact is true. She existed. She wrote. She died. What it added was not a fact. It was motion: a smile no camera caught, a glance no camera recorded, the first stroke of a pen that no one was there to film. And motion of that kind slips past the part of you that checks, because it never makes a claim. It just arrives, as feeling, and feeling does not ask for a source.
This is the wager, and you are allowed to think it is wrong. The greater danger is not the fake that lies to you. It is the true-enough thing that touches you. Thin history that feels intimate is more corrosive than fake history that feels off, because the counterfeit gets stopped at the border by a question, and the intimate one is waved through before any question is asked.
Two easier explanations are worth clearing first, because if either one is the real story, mine is not needed.
The first: the whole problem is that people take the clip for archival footage. But look at the ones who loved it. They were not fooled. One commenter wrote that it stirred them "to see her in 'real life,'" and the quotation marks around real life are doing quiet, honest work. The commenter knows it is not real life, and is moved anyway. So deception is not the mechanism. The feeling does not need the lie. It survives knowing.
The second: it is simply tasteless, a desecration. Maybe. One viewer answered the clip with two flat words, "Disgusting ai," and said the girl "deserves better." But the daughter of a survivor called the very same clip "so incredible." The same moving face pulled some toward it and turned others away. If bad taste were the harm, the harm would not land on the moved and the repelled alike. Distaste is a verdict on the clip. It is not what the clip does to your grip on the real record. So set taste aside too.
What is left is the mechanism. The felt sense of contact arrives first, and it arrives whole, and it is very hard to un-feel once the slower mind catches up.
We have known this about pictures for a while, in rooms that have nothing to do with grief. Show people a claim with a photo beside it that proves nothing, just a picture of the thing the claim is about, and they rate the claim likelier to be true. The picture is not evidence. It only makes the claim feel checkable, and feeling checkable turns out to be enough. In one study, a doctored childhood photo of a balloon ride that never happened led about a third of people to "remember" the ride by the end of one sitting, and half of them after a few. Show people doctored images of a public event and they recall a bigger crowd, more violence, more confrontation and injury than the day actually held. The image does not have to prove the memory. It only has to supply the feeling of one, and the mind fills the rest and files it under things I know.
Then there is a second, colder finding, the one that turns a worry into a thesis. When people are shown convincing synthetic video, the most common result is not that they swallow the lie. It is that they grow unsure of everything. Uncertainty climbs, and that uncertainty lowers their trust in the news on their own feeds. Follow that one step past what the study measured. Once made motion is everywhere, the real twenty-second film of a girl at a window may start to feel like just one more clip, no more solid than the desk and the pen. A board member of the foundation tied to her legacy put the endpoint plainly: in ten years, he warned, you will not know what is true and what is invented, which footage is real and which was generated. The fake does not have to win. It only has to blur the line until the real thing can no longer prove itself either.
Now watch how ordinary this gets. This year the tools left the studios. Your phone will animate a still photo of your grandmother in a few seconds: a wave, a party hat, confetti, a small heart. This month a photo app began relighting and restyling your old videos in a few taps, a morning glow, a wash of watercolor. It is sweet. It is nothing. No one is harmed when a birthday snapshot waves.
But the gesture does not change when the photograph does. The same wave, the same few taps, reach the face of a child who did not live to grow up. The software does not know the difference between a party and a grave. For a half-second, before the mind catches up, neither do you. That is the whole distance the technology has left to travel: none. The button that makes your uncle wave is the button that makes a murdered child smile.
This is where the argument has to turn, because there is a strong case on the other side, and I feel its pull.
The witnesses are dying. About 220,000 survivors were alive last year, median age eighty-seven. Within ten years, seven in ten of them will be gone; within fifteen, nearly all. When they are gone, there is no one left who was there. And a recording no one watches preserves nothing: attention has moved to screens that reward motion and answer back. A 2020 survey of American millennials and Gen Z found that nearly two-thirds did not know six million Jews were murdered, and half could not name a single camp or ghetto. Into that gap, the careful builders argue, a moving witness is not a desecration. It is a lifeline.
And the careful ones really are careful. One testimony project keeps a survivor to the only two languages she actually speaks, German and English, and refuses to let the machine make her seem to answer in any other. The reason is exact: if she ever appeared to speak a language she never knew, someone later could call the whole interview fake, and the tool would have destroyed her testimony instead of saving it. The constraint is the respect. In another project, a survivor who spent the war in hiding sat under the lights and answered some four hundred and fifty questions, and the system that will speak for her draws only from those recorded answers and reaches for nothing she did not say. She joked that she would be dead and still talking, in her red pants. "It's a responsibility," she said. "I just have to do it." An exhibition of pre-war faces goes a step further and prints the original photograph on the back of each reconstructed one, so you can turn it over and check the machine against the record. That is not a company mining grief. That is a witness agreeing to keep answering after she is gone, so that someone long after can ask her a question and hear her own recorded voice.
Put at its strongest, the case is this: intimacy can be a bridge to the evidence rather than a substitute for it. The pull toward a moving face can be aimed. It can be aimed at the recorded voice, at the original photo on the back, at the caption that leads to the source. Even one of the people most alarmed by all of this, who in the same breath calls the technology "a convenient tool for Holocaust distortion," will tell you that refusing to use any of these tools only leaves memory further behind. And that much is true.
The line is not simply real against fake. Both the clip and the bounded projects put a machine between you and a witness or a face from history, and both use it to hold a fading person in reach. But they are not doing the same thing. The bounded projects retrieve recorded testimony, or point back to an original photograph, and send you toward the source. The clip supplies motion that was never recorded at all. The difference is a single direction: toward the evidence, or away from it.
The bounded projects spend the feeling to send you to the record. There is her real recorded voice. There is the original photograph; turn it over. There is the question you can ask, and the true answer you will get. The intimacy is a door, and behind the door there is a source.
The true record is poorer than the fake. Most of us, honest about it, would reach for the fake. That preference is the whole danger.
The clip spends the feeling and, for the part that moved you, gives you nowhere to go. The caption is true and the diary is real. But the thing that reached you was the motion, and the motion has no source behind it: no camera ever recorded it. There, the door is painted on a wall. And it works anyway. It drew sixteen thousand likes and some comments saying she felt more alive, because the fake is warmer than the real one. Set the two side by side and the made motion wins. The desk and the pen pull harder than the real film, the twenty dull seconds of a girl leaning out a window at a wedding, showing almost nothing. The true record is poorer than the fake. Most of us, honest about it, would reach for the fake. That preference is the whole danger, and no label fixes it, because a label is a fact and the preference is a feeling, and the feeling arrives first.
Think about what the most careful people in this story are actually doing. They are building a survivor a voice that will outlast her, and then spending most of their care making sure no one can ever call that voice fake. The task of preserving a witness now includes protecting her from the very tools we would use to preserve her. That is where we are.
So go back to the window. Twenty seconds, Amsterdam, July 1941. A girl leans out to watch a neighbor's wedding go by in the street. She is not writing. She does not know a camera has caught her; it is there to film the wedding, and it finds her by chance. The film is barely twenty seconds, and that is everything the world was ever given of her, moving. We would trade it without pausing for the desk and the pen, and that trade is the one we are now free to make, a thousand times a day, on any face we have lost.
The next time a dead face moves and something in you leans toward it before you can think, you will have one question left that matters, and it will be the smallest one, and the least welcome. Not is it moving. Who saw it.