October 20, 2026

AI Cannot Act

A photorealistic “AI actress” had Hollywood’s agents curious and its actors furious. The debate is usually framed as a question of technology. It is really a question of what acting is.

In September 2025, at a film festival in Zurich, a young woman with an easy smile and flawless skin was presented to the entertainment industry as its next star. She had no stack of rejection letters, no cold-reading nerves, no first role in a school play. She also had no pulse. Tilly Norwood, the creation of the Dutch technologist Eline van der Velden, was generated by software, and within days the American actors’ union SAG-AFTRA had declared that she was not an actor at all.

Her maker called her “a piece of art”. Her critics saw a composite stitched together from the uncredited work of real performers. Both camps were circling the same unexamined assumption: that the only thing separating a synthetic performer from a human one is the quality of the rendering. Wait for the next model, the logic goes, and the gap closes.

It will not close, because the gap was never in the pixels. A model can reproduce what a performance looks like. It cannot do what a performer does: stand in a room with another person and decide, in real time and at real risk, what to give them. Acting is not the manufacture of an emotional surface. It is an act, a deed committed by someone who could have done otherwise and who answers for what they did. What follows makes the case that no system, however fluent, performs that deed, and that the industry’s own contracts, lawmakers and audiences are quietly coming to the same conclusion.

A star with no pulse

The outrage that greeted Tilly Norwood overlooked how little autonomy she had. Her first screen appearance, a comedy sketch called AI Commissioner released in July 2025, was built with ten AI tools and a script written by ChatGPT, then shaped by a human team. Her studio, Particle6, has claimed that using her could cut production costs by as much as 90 per cent. Van der Velden’s pitch to the industry was candid about the economics: budgets are shrinking, demand for content is rising, and audiences, she wrote, care about the story, not whether the star has “a pulse.”

The reaction was swift. SAG-AFTRA said Norwood was a character generated by a program trained on the work of countless performers without permission or compensation. Britain’s Equity, which represents nearly 50,000 performers, warned of an AI “Wild West.” Emily Blunt called the prospect terrifying. Whoopi Goldberg noted that a human is suddenly up against something generated with thousands of other actors. Mara Wilson asked about the living young women whose faces, she suggested, were composited to make her.

A year on, Norwood is still with us. In July 2026 Particle6 announced her first feature, Misaligned, a comedy-drama set in a cloud-based “Tillyverse.” Its heroine, as Variety describes her, is an AI being with access to everyone else’s life experience and none of her own. Read that again. The first feature built around the most famous synthetic performer in the world takes, as its central premise, the very limitation at the heart of this argument.

What an actor actually does

Ask a great acting teacher what the job is and the answer is rarely “emote.” Sanford Meisner, the American teacher whose technique is among the foremost in modern actor training, built his technique on the opposite instinct: stop watching yourself and attend to the person across from you. In his best-known drill, two actors repeat a single phrase to each other until the words wear thin and what remains is behavior: a flicker of impatience, a softening, a change in the weather between two people. The exercise removes line readings altogether, because the sound of each line is shaped by what the partner just did. Meisner’s craft is often summed up as “living truthfully under imaginary circumstances.”

Every word of that sentence is a problem for a machine. Living: the performer has a body that can be moved by what it hears, a pulse that quickens, a breath that catches. Truthfully: the response has to be produced by this moment, not selected because it resembles responses that worked before. Imaginary circumstances: the actor bridges the invented situation with a real store of memory, a particular loss or humiliation or joy. That is not information. It is experience, and it is why two actors given the same line deliver it differently.

SAG-AFTRA’s own teaching note on Meisner puts the emphasis on listening, not to the words but to the other actor’s behavior. Coach Alex Taylor tells Backstage that a present, listening actor gives an editor a real person to cut to, rather than a prop. Whatever happens inside a model, it is not in the room, it has nothing at stake in the scene, and it cannot be changed by the person opposite it the way an actor is. It can produce lines. It cannot be surprised into a response.

The gap in the machine

Start with the stage, because it is the closest thing we have to a controlled experiment. In 2023, researchers from the University of Kent and the improv troupe Improbotics put a language-model “actor” into 26 live shows at the Edinburgh Festival Fringe, playing scenes alongside human improvisers. A human operator fed it context and picked its best lines; a human performer spoke them. Audiences were generous: more respondents said they were rooting for the AI than anything else. Asked to rate it as a performer, though, they scored it low on communicating naturally and high on being oblivious to what was happening in the scene. The human cast members had to justify its non sequiturs and steer the plot around it, which is to say they acted on its behalf, and one noted that its best lines were elevated by the person delivering them.

The authors caution that this was an informal survey, around 150 audience responses collected between shows at a crowded festival, using 2023-era models. Models have improved. But the finding is less about scores than about what was missing: the ability to track a relationship, hold a thread across a scene, and take a risk for a partner.

Now look at the best-resourced synthetic performer in the world. When Particle6 released a music video for Norwood in March 2026, the studio’s own press release explained that a new technique, performance capture, meant van der Velden herself had acted out Norwood’s performance, and that the team behind the video comprised 18 real humans. Her creator’s summary: “people remain at the heart of it.” Reviews of the earlier AI Commissioner sketch, as summarized here, were unkind: The Guardian found it technically competent but relentlessly unfunny, with stiff delivery, and The A.V. Club reported a strong uncanny-valley effect.

A deeper problem is that audiences do not respond to a performance in a vacuum. In a study in Scientific Reports, scientists at the Science of Intelligence research cluster recorded ratings and EEG responses from 30 participants shown smiling, angry and neutral faces they were told were real or computer-generated. When faces are otherwise indistinguishable, the researchers concluded, responses may depend crucially on the prior belief about whether the face is real, and knowing a face is artificial may blunt its impact, especially for positive emotions.

Honesty requires one concession. In blind tests, machines are becoming hard to catch. A Crowd React Media study in May and June 2026 played promos voiced by a professional or an AI to 1,326 U.S. radio listeners, who rated the two almost identically on professionalism, authenticity, credibility, energy and likability. The one statistically significant gap was humor: 33 per cent called the human delivery funny, against 26 per cent for the AI version. Opinions of the AI voice shifted, significantly, once listeners were told what they had heard. The surface of a performance is becoming reproducible. If the experience changes the moment we learn nobody is there, then the nobody was part of the experience all along.

The contract and the law

If a machine could simply act, the people who sign the cheques would not be bargaining over it so intently. AI was a sticking point in the 2023 actors’ strike, and this summer’s sequel went further. Members of SAG-AFTRA ratified the 2026 TV/Theatrical Agreement on June 4 by 91.42 per cent to 8.58 per cent. It runs from July 1, 2026 to June 30, 2030.

For “synthetics,” the term for fully generated performers like Norwood, producers may not proceed without first bargaining with the union on a strict notice schedule, with arbitration and damages if they breach it. They have agreed to a principle favoring human performances and committed not to put a synthetic in a role a human would otherwise play unless it brings “significant additional value” over a human or that human’s digital replica. Replicas of real performers, the union’s chief negotiator told AFP, must have informed consent and fair compensation.

This is not a ban, and IndieWire’s analysis was careful to say so: nothing in it stops a studio from using a synthetic if it chooses. What it does is reverse the burden of proof. The studio must justify the machine; the actor no longer has to justify the human.

Industry and Congress are circling the same idea. In October 2025, Bryan Cranston found that his voice and likeness could be generated in OpenAI’s Sora 2 despite his opt-out. After talks with SAG-AFTRA and the major talent agencies, OpenAI strengthened its guardrails and joined them in backing the NO FAKES Act. That bill, S. 4591, would create a federal right over a person’s voice and visual likeness. It cleared the Senate Judiciary Committee unanimously on June 18, with an unusually broad list of backers that includes SAG-AFTRA, the Motion Picture Association, Disney and OpenAI. Critics such as NetChoice warn its takedown rules could push platforms to remove lawful content. As of this month it was still awaiting a Senate floor vote.

Read these instruments together and notice what they protect: people. Their consent, their pay, their faces, their voices. In each, the machine’s output is treated as derived from human performance, never as a performance in its own right. The law is quietly adopting the position of this essay, that a performance is the act of someone, and that someone is owed something for it.

Why audiences still need a person

Every great performance carries a quiet piece of information: someone did this. A tightrope walker’s feat matters because the walker can fall. A rendered figure cannot fall, and no rendering, however exact, restores the stakes. The same is true, in subtler form, of a face crumpling on screen. We read it as the product of a real person’s willingness to feel something in front of strangers. SAG-AFTRA’s contention is that audiences have little appetite for computer-generated content cut loose from human experience, and the studio behind Norwood argues the opposite, that viewers care about story, not pulse. The honest position is that the market has not yet settled the question. But the early evidence leans one way.

The Edinburgh audiences were asked what they liked best about the AI-assisted show. Forty of the 150 respondents named watching the robot and the humans make stories together as the best part; 37 per cent left more excited about using AI tools for creativity, against 16 per cent who said they were more optimistic about AI as a storyteller, according to the study’s authors. People liked the machine as an instrument in human hands. They were less sure about it as the author of a scene.

Meanwhile, the most inventive responses to Tilly Norwood have been human performances of her. A TikTok creator went viral with skits in which she plays Norwood auditioning for roles and failing, on her first day on set, to follow simple instructions. A mobile-phone company’s ad enlisted a real woman named Tilly Norwood to mock the character. The synthetic star became a role that people played, which is what stars have always been.

What AI is actually good for

None of this makes AI useless to performers. It already de-ages faces, dubs shows into other languages and polishes visual effects, as CBC has noted, and it was used to alter voices in two best-picture nominees that went on to win multiple Oscars, to a mixed public reaction. The 2026 SAG-AFTRA agreement has dedicated terms for exactly these uses: digital replicas, dubbing and language consent, biometric data, and alterations that stay within what the work originally scripted. The logic is consistent. The technology can extend a performer’s reach into new languages, ages and stunts. It does not originate the act.

Even the preferred metaphor of Norwood’s creator makes the case. She has described AI as “a new paintbrush." If so, the work belongs to the painter, and nobody credits the brush. By the logic of her own analogy, which also invokes animation and puppetry, Norwood is a character realized by artists, and animated characters have never needed to be actors; the people who voice and animate them are. Judge the work as animation if you like. It is the claim to the title “actress” that fails.

The Edinburgh researchers reached a similar view from the other direction, describing their own show as closer to human augmentation than to autonomous AI. That may be the most defensible future for the technology on a stage or a set: a powerful instrument that performers pick up, with consent and credit, and put down.

Curtain call

The most revealing thing about Tilly Norwood is her own premise. Her first feature casts her as a being with access to everyone else’s experience and none of her own. That is not a flaw in the casting. It is a precise description of what these systems are: instruments trained on the accumulated record of human performance, able to produce its surface at any hour for almost no cost, and unable to stand behind a single choice of their own.

Acting is the opposite. It is one life made visible to others, in real time, with something at stake. That is not a forecast that the technology will stop improving; it will not. Studios will keep experimenting, some of the experiments will be good, unions will keep bargaining and Congress will keep arguing. Pixels can be perfected. But the question was never whether a machine can look as if it is feeling something. It is whether anyone is there.

In acting, someone has to be.

Watch and look

Video and listening, with who made each and why it matters here:

Title Who Why it matters
AI Commissioner Particle6 The 2025 sketch that introduced Tilly Norwood; judge the performance for yourself
Take the Lead Particle6 The March 2026 music video built with performance capture by a human actor
Sanford Meisner master class The Sanford Meisner Center Footage of the teacher on listening and truthful response
Unpacking the 2026 TV/Theatrical Agreement SAG-AFTRA (podcast) The union’s leaders explain the AI terms in their own words
SAG-AFTRA on YouTube SAG-AFTRA The union’s video channel

SAG-AFTRA’s Synthetic Performers explainer graphic summarizes the 2026 contract terms on a single page.

Sources

Accessed October 20, 2026. Items marked † could not be opened in full; claims drawn from them rest on search excerpts.

The synthetic performer

What acting is

Research

Contracts and law