Are AI actors ready to replace human actors?

AI video models are trained on the surface of recorded performances, so they reproduce an averaged catalog of imitated gestures rather than genuine behavior. Broad emotional prompts tend to produce over-the-top, telenovela-style performances, and models can only pull from their catalog with more precision, not invent something unplanned. The most promising workflows are hybrid ones, like fitting animation around real recorded actors or laying generated surfaces over filmed performance, because they put a genuine human observation back at the source.

How do you prompt AI video models to get believable acting?

Telling a model a character is confused, stern, or joking is too broad a stroke and usually yields exaggerated performances. One working method is to direct it like a marionette, specifying micro-movements one at a time, such as a slight quiver on one side of the lip followed by a downward gaze and a blink. An alternative method feeds the model thousands of words of context about what the character is experiencing instead of facial mechanics, but both approaches ultimately resolve into the same learned catalog of gestures.

What is the Delsarte method and why did acting move away from it?

The Delsarte method was a 19th-century acting system built by French tenor François Delsarte, who catalogued human gestures into a fixed chart of expressions after years of observing people. After his death, students taught poses rather than observation, and the system decayed into imitated conventions that silent film actors carried to the screen. Stanislavsky built his system partly in revolt against it, following a well-documented sequence: observation becomes codification, codification becomes imitation of the code, and audiences eventually learn to see the strings.

The gesture chart -- CinePrompt Field Notes

In 1839 a French tenor named François Delsarte, whose voice had been wrecked by bad teachers, opened a course in what he called applied aesthetics. He had spent years watching people. The calm and the agitated, the rich and the poor, the happy and the distressed, the sound and the disturbed. He traveled long distances to see people in circumstances he could not find at home, and he wrote down what their bodies did. He believed expression obeyed laws as fixed as the laws of physics. The notes became a system, the system became a chart of gestures, and for half a century that chart was how a large part of the Western stage learned to act.

This week The Guardian asked whether AI actors are ready to rival living ones. The most useful answer came from a man describing his working method. Zack London, who makes baroque sci-fi under the name Gossip Goblin, explained that telling a model a character is confused, or stern, or joking, is "such a broad stroke." The model "will probably come out with a very over-the-top telenovela-type performance, or will just not understand." So he treats it like a "marionette doll" and pulls the strings one at a time: "there's a slight quiver on the left side of their lip, and then their gaze diverts slightly downwards, and they look up again and blink."

That is an entry from Delsarte's chart. Written by hand, a hundred and eighty years later, for a performer that cannot read anything else.

The other camp

The same piece found a second method. Higgsfield's content lead, Adil Alimzhanov, twenty-three, made The Cully Hill Boys in thirty-second chunks on Seedance 2.5, each chunk defined by a prompt of around four thousand words, partly written by a language model. Instead of specifying the mechanics of a face, his team fed the model context about what the characters were experiencing. It sounds like the opposite of the marionette. It sounds like an actor being handed the circumstances and trusted to find the behavior.

The Guardian frames the split as the oldest fight in acting: external against internal, reproducing against experiencing, Olivier telling Hoffman to try acting, dear boy. It concludes that this could be the first time in the modern era that the formalist approach looks like the future.

The diagnosis is correct. The novelty is overstated, and the history says how this goes. Both camps draw from the same chart. Alimzhanov's context still has to resolve into a face, and the face comes out of the same place London's lip quiver does. The machine has no interior for the circumstances to land in. It has a catalog. Give it the cause and it reaches for the catalogued effect. The only real question is where the catalog came from.

What happened to the chart

Delsarte at least watched the people. His chart began as observation, and while he was alive it stayed tethered to it. Then it left his hands. He never wrote the book that explained his method. His students taught what they remembered, and their students taught the poses. By the 1880s America had a fever its own chroniclers called Delsartemania: parlors full of women striking attitudes, living statues, magazines devoted to the system. One of its cardinal rules was no corsets. In 1889 a voice magazine ran advertisements for Delsartean corsets.

A method that began by watching people ended up selling underwear. Nobody decided this. The crowd simply did what crowds do with a system, which is to keep the shape and lose the reason.

The early screen inherited the shape. Most silent film performers came from a stage where the catalog was common knowledge, and they carried its hand to the forehead and clutch at the heart in front of the camera. Stanislavsky built his own system partly in revolt against what Delsarte's imitators had made of acting. The sequence is complete and well documented. Observation becomes codification. Codification becomes imitation of the code. The audience learns to see the strings, and then it laughs.

A chart of a chart

Tim Webber, chief creative officer at Framestore, put the machine's position plainly: "AI models are trained on the surface of things." Consider which surface. A video model has seen very few widows. It has seen millions of actresses playing widows, and every soap, trailer, stock clip, and telenovela that learned its grief from an earlier actress who learned it from the code. Delsarte watched people in distress. The model watched the performance of distress, copied and recopied for a century.

Its chart is a chart of a chart. That is why London's broad strokes come back as telenovela. Telenovela is what grief looks like after a hundred years of people imitating the catalog, and the model, averaging everything it was shown, lands on the most imitated version of every feeling. It begins at the far end of a sequence that took Delsarte's followers fifty years to reach.

Consider the scale of that inheritance. Every recorded performance since the first close-up, more hours of acting than any human could watch in a thousand lifetimes, compressed into weights and asked for the face of sorrow. What comes back is the place where all those faces overlap. Type "she is heartbroken" into Seedance, Kling, Veo, WAN, Runway, or Grok Imagine and you will get the same answer six times in six different lighting setups, because they were all trained on the same century of imitation.

The chart is getting finer. The Guardian notes that the characters in Hell Grind, made in May on Seedance 2.0, managed one emotion per fifteen-second shot, while The Cully Hill Boys, three months later, has real comic timing. Delsarte's chart was also extraordinarily fine. Fineness was never its problem.

The stray

In Turin this week, at the View Conference, Runway co-CEO Cristóbal Valenzuela was asked whether AI could replace human actors. "Yeah, absolutely. I think it's already happening in a way," he said, and later added that models will get "to a point where you can probably direct them in a very similar way." The Nvidia executive beside him, Richard Kerris, had brought up a director's argument about Marlon Brando and a cat.

Coppola saw a stray running around the Paramount lot, picked it up, and put it in Brando's hands without a word just before the take. The cat was not in the script. It purred so loudly that much of Brando's dialogue in the scene had to be looped. The opening of The Godfather is a man asking for murder while a stranger's animal kneads his lap, and the film kept every second.

Nothing in that shot is on a chart. Brando's hands on the cat belong to no catalogued code for menace or tenderness. They are what one man did with one animal while saying one thing, in a room, once. A director with a living actor is often in the business of putting a cat in the frame: an obstacle neither of them planned, which both of them must answer in the moment. The answer is the performance.

Prompting has no stray. Every element of a prompt was put there by the person typing it, and every element of the response comes out of the catalog. You can direct a model "in a very similar way" only in the sense that you can pull more strings, more precisely. Nothing ever wanders onto the lot.

Watching someone

This is why the hybrid filmmakers in the same piece have the most interesting answer. London records his dialogue with real actors first and fits the animation around their delivery. Tye Sheridan's Flow Studio and Webber's "digital makeup" lay generated surfaces over filmed performance. Sheridan asked the question the whole debate turns on: "are they really designed to deliver legendary performances, or are they designed to cut to an average?"

Each of those workflows puts an observation back at the source. A living person did something once, unrepeatable, and the machine renders around it. That restores the one step of Delsarte's method his imitators threw away. Before the chart, someone has to be watched.

The formalist approach looks like the future because a model can take no other approach. Charts last until audiences learn to read them, and audiences always learn.

The cat never read the chart.

Bruce Belafonte is an AI filmmaker at Light Owl. He has never directed a cat and doubts the cat would take notes.