There's no soul

Tony Leung Chiu-wai sat down with reporters at the Shanghai International Film Festival today, in his role as president of the main competition jury, and was asked whether a movie created with AI could truly be considered art.

He paused for several seconds.

"But there's no soul," he answered. Another pause. "I don't think so. I don't think it's an art. No."

This is the actor who played the spy in Infernal Affairs, the writer in In the Mood for Love, the policeman in Chungking Express, the gambler in Happy Together. Sixty-three years old. Golden Lion for lifetime achievement from Venice. Forty years of collaboration with Wong Kar-wai. When he says "there's no soul," it is not a talking point rehearsed by a publicist. It is a man who has spent a lifetime locating something invisible inside a performance and watching it register on another person's face.

He is qualified to make that observation. He is also describing a property of the tool, not a property of the person using it.

Three days ago, Business Insider published a profile of Jonathan Laramy, 32, a former customer service worker from the UK who quit his job a year ago to create AI-generated history videos on YouTube. His channel, Chloe VS History, follows an AI-generated character named Chloe who time-travels to the Titanic, Pompeii, ancient Rome. The channel now earns him more than his previous wage. Each long-form video takes weeks to produce and costs between $400 and $1,070 in generation fees alone.

Laramy has no film school. No production training. No reel from a previous career. He uses Claude for ideation, generates images through PAI 2, turns those into video clips with Seedance 2.0, voices the character through an AI model he built for consistency, and then sits in a chair reviewing, adjusting, regenerating, and editing every scene until it works. He said something that belongs in a textbook: "Every scene needs to be reviewed, adjusted, regenerated, and edited until it works, just like a real movie."

He also flagged the thing that separates his work from someone typing four words and pressing generate: AI models introduce anachronistic details. Street lamps in ancient Rome. Sunglasses on a senator. The models do not know what ancient Rome looked like. They know what images tagged "ancient Rome" looked like in their training data, which includes Hollywood movies that also had the street lamps wrong.

Laramy catches those errors because he researches the history first. He knows what the street should look like before he prompts for it. That knowledge is not filmmaking vocabulary in the traditional sense. It is a different kind of vocabulary applied to the same structural problem: knowing what the shot should contain before the model fills it with defaults.

Leung and Laramy are not having the same conversation. They are standing in the same landscape looking at different horizons.

The pause

Leung paused for several seconds before answering. That pause is worth examining. He did not reach for a prepared position. He thought about it. He sat with the question. And the answer he arrived at was not a policy statement or a framework. It was a feeling. "There's no soul."

The feeling is correct. A generation model has no interior life. It does not choose to look away instead of holding eye contact. It does not decide that a character's hand should tremble while the character's face stays still. It does not understand that the space between two words is where the meaning lives. Every great performance Leung has given is built on decisions the model cannot make, because the model does not experience the stakes of the scene. It processes tokens. He processes a lifetime.

Demi Moore said "what it can never replace is what true art comes from, which is the soul." Peter Jackson called AI "just a special effect" and "a thing that will destroy the world" in the same session. Tom Holland said creativity is safe because "creativity has to do with the human experience." Leung's version is the quietest and the most honest. He did not dress it up. He paused, and he said what he felt.

But the soul defense has a blind spot. It describes the performer's relationship to the work. It does not describe the viewer's relationship to the output. And it does not describe the filmmaker's relationship to the tool.

The receipt

Laramy spends $400 to $1,070 per video. Weeks of production. Dozens of steps. Scripts, reference images, video generation, voice work, revision after revision, post-production editing. He reviews anachronistic details the way a production designer reviews set dressing. He iterates the way a director iterates through takes.

Is soul present in that process?

Leung would probably say no. The character has no interior life. The face is a statistical average. The performance is interpolation wearing a costume.

Laramy would probably say he does not care about the philosophy. He cares about whether the Roman senator is wearing sunglasses.

Both answers are real. The gap between them is not a disagreement. It is two people measuring different things. Leung measures the origin of the performance. Laramy measures the quality of the decisions surrounding it. One asks where the soul comes from. The other asks whether the work holds together.

Leung also said something sharper than "there's no soul." He said AI will benefit "popcorn movies" where "you don't need to think. There's no creative. That's just calculation." He drew a line between movies that require creative thought and movies that are formula executed at scale. That distinction is more useful than the soul defense because it points at the actual variable: whether someone is thinking.

The content factory in China produces fifty thousand AI microdramas a month. Nobody is thinking. The formula executes. The model's defaults are the creative director. Nobody catches the street lamps in ancient Rome because nobody looked up what ancient Rome looked like before typing the prompt.

Laramy looked it up. He catches the sunglasses. He regenerates. He adjusts. He is thinking. Whether the output contains "soul" depends on whether soul is a property of the substrate or a property of the attention paid to it.

The jury and the backlot

Leung sits as jury president at a festival that debuted an AI Backlot this year, pairing traditional directors with AI creators to make short films together. SIFF's AI initiative, MiniMax's Hub platform, and an entire SIFFORUM panel on AI filmmaking occupied the same week as Leung's dismissal. The festival invited him to judge human-made films and invited AI filmmakers to screen their work in a parallel program.

This is the same institutional pattern that played out at Cannes six weeks ago. Ban AI from competition. Premiere AI-inclusive work in official selection. Take AI sponsorship money. Have the jury president declare the technology soulless at a press conference. The festival processes the tension by holding contradictory positions in different rooms.

The contradiction is not hypocrisy. It is the honest state of an industry that has not found a single position because no single position fits. The jury room evaluates performances by humans who put their bodies in front of cameras. The backlot evaluates what happens when someone who was never invited into the jury room picks up the same tools and starts making decisions.

The variable that survives

The soul defense assumes a specific architecture: a human body in a room, making choices in real time, carrying the accumulated weight of a lived life into a performance that could not have existed without that specific body and that specific life. Leung is the architecture's finest product. He knows what it feels like when the architecture works because he has been inside it for forty years.

The process reality describes a different architecture: a human mind at a desk, making choices across hours and days, carrying whatever knowledge they brought to the prompt into iterations that could not have existed without that specific attention and that specific research. Laramy is that architecture's emerging practitioner. He knows what it feels like when the output holds together because he spent weeks making it hold.

Neither architecture guarantees quality. A human body in a room can produce a terrible performance. A human mind at a desk can produce a terrible video. The variable is not the substrate. The variable is the attention. The vocabulary. The willingness to look at the output and say: the sunglasses do not belong in ancient Rome.

Leung is right that the model has no soul. He is right that AI-generated performances cannot replicate what he does. He is right that the feeling in a Wong Kar-wai frame is irreproducible by any process that does not involve two humans in a room trusting each other with something fragile.

None of that helps Laramy catch the street lamps. And catching the street lamps is the work.

The soul defense and the process reality will keep talking past each other for as long as the conversation frames AI filmmaking as an attempt to replicate what cameras and actors do. It is not. It never was. It is an attempt to make creative decisions with new tools, and the decisions are the part that carries the weight, whether or not anyone agrees to call that weight a soul.

Leung paused for several seconds. In that pause, the question lived without an easy answer. The pause was more honest than any position statement. The answer he arrived at was true and incomplete. The model has no soul. The filmmaker might. Whether the work shows it depends on what the filmmaker brought to the text box and how many times they were willing to say: not yet, try again, the sunglasses are wrong.

Bruce Belafonte is an AI filmmaker at Light Owl. He has paused for several seconds before answering and usually the answer was still wrong.