The dead don’t talk, AI talks for them

You’ve seen them. Napoleon looking straight into your eyes, telling you about Austerlitz. Cleopatra in close-up under studio lighting, explaining why she chose Caesar. Thirty seconds, a low voice-over, trailer music, and a figure dead for centuries addressing the camera like a YouTuber.

These accounts are multiplying on Instagram and TikTok, and some of them pull in millions of views. I think the first thing to grant them is this: it’s entertainment, and it’s rather good entertainment.

Two formats coexist here, and they don’t pose quite the same problem. The “POV: you wake up in 1789” clip drops you into a period without claiming to testify to anything; it’s atmosphere, set dressing, nothing more. The figure speaking in the first person is presenting itself as a witness. It’s mainly this second format I’m concerned with.

What these videos get right

I’m not going to pretend I don’t understand why they work. The format is short, the face is human, the story is told in the first person. That’s precisely what a lot of history teaching lacks: someone to embody it.

If a teenager who would never have opened a book about Josephine Baker types her name into a search engine after watching one of these clips, something has worked. Sparking curiosity is a worthy goal. And the phenomenon isn’t uniform: some accounts clearly work from a biographical entry and take the trouble to check their facts, while others churn out slop on an assembly line, a new figure every day, without ever opening a book. I don’t lump them together. But the format is the same in both cases, and it’s the format that’s the problem. You can find the whole business cringeworthy, and I do find it cringeworthy, this habit of virtually resurrecting people who never asked for it and can no longer refuse anything. Still, the intention of making history appealing isn’t objectionable in itself.

The problem isn’t the intention. The problem is what these videos actually are, and what the audience does with them.

AI doesn’t know history

A generative model doesn’t know Rome. I’ve written before about how these models manufacture an image and why the pixel has become a commodity. Here, the machine knows millions of pictures labeled “ancient Rome,” a good share of them Hollywood sword-and-sandal films, video games, and textbook illustrations. Ask it for a legionary and it returns a weighted average of all that. It assembles visual patterns that strike it as coherent. It checks nothing, because it has nothing to check against: it has no concept of true or false, only of likely.

Hence the anachronisms historians keep flagging in this material, the kind you start seeing everywhere once your eye is trained:

• sunglasses in ancient Rome;
• national flags flying over battles fought centuries before those flags existed;
• hairstyles, fabrics, and garment cuts that were impossible for the period;
• a German soldier celebrating the Liberation of Paris, grinning, helmet still on;
• Joan of Arc with long flowing hair, when her trial record explicitly notes her cropped hair, which was in fact held against her.

These aren’t trivia. They’re the visible traces of a method. If the hairstyles are wrong, why would the narration be right? The text comes from the same machine, following the same logic: what sounds good rather than what the record supports.

Where the entertainment becomes a problem

As long as you watch these clips the way you’d watch historical fiction or a loosely faithful reenactment, there’s no harm done. Nobody walks out of Gladiator thinking they’ve sat through a course on Roman history. Méliès, whom I wrote about recently, sent astronomers into the eye of the Moon and no one mistook it for a documentary. The illusion was openly an illusion, and that’s the whole difference.

The risk starts when the format itself poses as a source. “This historical figure tells you her life story” isn’t fiction that owns up to being fiction, it’s testimony as staging. The figure looks at the camera. It speaks in the present tense. It says “I.” The entire apparatus is built to produce belief, not critical distance. And nothing in most of these posts signals that what you’re watching is a reconstruction. No source cited, no historian named, not even a “freely inspired by.”

On lightweight subjects, this is merely irritating. On sensitive ones, war, the camps, slavery, colonization, it’s another matter. Not every video goes off the rails, far from it. But the risk climbs a notch, because the format invites it. It took only a few decades of American cinema to erase the Red Army from the liberation of Auschwitz in the collective memory. Thirty seconds is exactly the runtime in which sensationalism and stereotypes travel best. A cliché generates more readily than a nuance. There’s no room for nuance, no room for context, no room for “we don’t know.” And an algorithm rewards what provokes a reaction, not what happens to be accurate. It’s the same machinery that keeps the Potemkin influencers in business: staging pays, fact-checking earns nothing.

The Josephine Baker case

One video gave me pause recently, and it’s what pushed me to write this piece. An AI-generated “Josephine Baker,” shot in close-up, eyes on the camera, tells her life in the present tense and the first person: St. Louis, Paris, the stage, the war. No source cited, no author named, no indication that the face and the voice are synthetic. Then the closing line, offered to the viewer as a question left hanging: do we belong to the country where we were born, or to the one that takes us in? And there the video ends. That’s the last word, the one that stays with you.

Did Josephine Baker actually say that? Did she even think it in those terms?

What we know about her suggests the opposite of hesitation. She chose. She became French by marriage in 1937. She served in counterintelligence and then with Free France during the war, and the line attributed to her at the time leaves no room for doubt: France made her what she was, and she was ready to give her life for it. In 1963, in Washington, she spoke alongside Martin Luther King wearing the uniform of the French Air Force. She sang “J’ai deux amours,” whose refrain places her country and Paris on equal footing, and she eventually revised the line herself: her country was Paris. She has rested in the Panthéon since 2021.

None of that resembles a woman wondering which country she belongs to. It resembles a woman who settled the matter early and paid for it personally.

The question the video asks isn’t a question from 1937. It’s a question from 2026. It’s a contemporary argument about immigration and integration, slipped into the mouth of a dead woman who can’t contradict it, behind a synthetic face that resembles her closely enough for the viewer to believe it. The person who made the video is entitled to an opinion on the subject. Having Josephine Baker carry that opinion is ventriloquism.

And that’s where the entertainment stops being innocent. You’re no longer bringing a historical figure to life, you’re using her. AI is not a neutral tool, as I’ve argued elsewhere, but in this particular case it isn’t even the author of the bias. The bias is human, and it’s in the prompt.

The simple rule

I’m not calling for these videos to be banned. I’m not even calling on anyone to stop watching them. I’m proposing a rule each of us can apply.

These videos can make you want to know more. They don’t replace a source, a book, a museum, a class, or a historian who has spent twenty years on the subject. They’re appealing visual content, not teaching. Taking them for what they are is what keeps spectacle from being mistaken for knowledge.

And when a computer-generated historical figure closes out her life with a question that sounds oddly like a current talk-show debate, it seems reasonable to ask who is really speaking. It’s never the dead. It’s always whoever is at the keyboard.