
AI Actors: What’s Actually Happening
AI & Future of Film 16th September 2026 Jon Dingle

By the Filmoria News Desk
SAG-AFTRA members picketing during the 2023 strike. Photo: Phil Roeder, via Wikimedia Commons (CC BY 2.0).
No, an AI has not been cast as the lead in a studio film. What has happened is quieter and more consequential: real performers are being scanned, de-aged, revoiced and recreated under a set of rules that changed again in 2026, while a handful of companies push synthetic performers that nobody in the industry will sign. Here is what is actually in place, what is marketing, and what is worth arguing about.
Every few months a story goes round that Hollywood has cast its first AI actor. It is usually a synthetic face with a publicist, and the casting turns out to be an announcement rather than a job. Meanwhile the technology that genuinely is reshaping screen performance — digital replicas, voice cloning, de-ageing, crowd generation — has been quietly working on films you have already seen, mostly without a headline.
This is a map of where things really stand in September 2026: what the terms mean, what the new union deal permits, where the UK sits, and which parts of the panic are justified.
Three Different Things Called “AI Actors”
The phrase covers at least three separate technologies with separate rules, and the confusion is doing real damage to the debate.
1. Digital replicas
A scanned, modelled or trained likeness of a real, living performer, used to generate images or audio of them. This is the category that covers de-ageing, body doubles for dangerous work, dialogue fixes and posthumous appearances. The performer exists; the question is consent and payment.
2. Synthetic performers
A character with no human behind it — a face assembled from a model rather than captured from a person. This is what Tilly Norwood is. Union agreements treat it as a distinct category precisely because there is no performer to consent or be paid.
3. Generative video characters
People who appear in clips produced by text-to-video systems. They are not cast, directed or hired; they are generated per shot and are frequently inconsistent between shots, which is why almost nothing made this way has held together at feature length.
Most arguments about “AI actors” collapse because one side is talking about de-ageing Tom Hanks and the other is talking about replacing extras with software.
![]()
Photo: Applearts123, via Wikimedia Commons (CC BY-SA 4.0)
Tilly Norwood and the Synthetic Performer That Nobody Signed
The most famous AI actor in the world is a good illustration of how thin the reality is beneath the coverage. Tilly Norwood was unveiled in 2025 by Xicoia, an AI talent studio spun out of the production company Particle6, founded by Eline Van der Velden. The pitch was that she would be represented like any other performer.
The response from the industry was close to unanimous rejection. SAG-AFTRA’s statement was blunt about the provenance problem, describing her as a character generated by a computer programme trained on the work of countless professional performers without permission or compensation. Equity joined the criticism in the UK. Working actors queued up to say they would not work opposite her.
What happened next is the part that gets less attention: not very much. By early 2026 Tilly still had no agency deal, despite her creator telling the trade press that agencies were interested. The company’s stated plan had shifted from conventional casting towards building a “Tillyverse” — a social-first universe of AI characters with narrative arcs and fan interaction — and it hired a former Amazon Prime Video social strategist to run it.
That pivot is the tell. Synthetic performers are currently a content play for platforms where consistency and craft matter less, not a casting play for film and television. The economics of a scripted drama do not improve much by removing the actor, because the actor was never the expensive part.
What the 2026 SAG-AFTRA Agreement Actually Says
The single most important development for performers this year is not a piece of software. It is a contract.
SAG-AFTRA members ratified a new TV/Theatrical agreement on 4 June 2026 with 91.42% voting in favour on a 19.25% turnout. It took effect on 1 July 2026 and runs to 30 June 2030. Alongside wage and residual increases and a pathway to merge the two pension plans, it tightens the AI terms won during the 2023 strike.
The union’s framing is that the new terms impose stringent rules to ensure synthetics remain the exception in the industry rather than the rule. National Executive Director Duncan Crabtree-Ireland put the purpose plainly: protecting the value of human performance and creativity. President Sean Astin described the deal as strengthening protections around artificial intelligence and digital identity.
The core architecture, carried forward and reinforced, is consent-based:
- Informed consent to create a digital replica, separately from consent to use it.
- Consent per project, so a scan taken for one film cannot quietly become a career’s worth of appearances.
- Minimum compensation when a replica is used, scaled to how extensively it is used.
- Restrictions and notice around synthetic performers, rather than a free hand to generate a cast.
Two caveats matter. First, this is a union agreement, not a law: it binds signatory productions, and a great deal of screen content — adverts, branded content, online video, some international production — sits outside it. Second, a 19.25% turnout on a deal this consequential tells you something about how much of the membership feels able to influence the outcome.
![]()
Photo: Thomas Wolf, www.foto-tw.de, via Wikimedia Commons (CC BY-SA 3.0)
The UK Picture: A Government U-Turn and a Possible Personality Right
British performers have been fighting a parallel battle with a different opponent: not studios, but copyright policy.
For most of 2025 the UK government’s preferred option was a broad text-and-data-mining exception that would have let AI developers train on copyrighted work unless rights holders opted out. Equity called it legalising theft. The creative industries mounted one of the loudest lobbying campaigns in recent memory.
In March 2026 the government stepped back, saying it would not reform copyright law until it was confident the reforms met its objectives, and — more interesting for performers — signalled it was exploring a new personality right covering unauthorised and unpaid use of a performer’s voice and likeness. Equity’s general secretary Paul W Fleming welcomed what he called a marked change of approach.
A personality right would matter more to a working British actor than any union clause, because it would apply whether or not the production was covered by an agreement. It does not exist yet. By August 2026 performers and musicians were campaigning jointly against unlicensed voice cloning, which suggests nobody is treating the policy shift as a finished job.
Where AI Performance Has Already Shipped
Strip out the announcements and a real production record remains — almost all of it replicas of real actors rather than synthetic ones.
De-ageing
Robert Zemeckis’s Here (2024) used Metaphysic’s system to de-age Tom Hanks and Robin Wright with results the actors could see on the monitor as they performed, rather than waiting months for a post-production reveal. That immediacy is the genuine shift: de-ageing stopped being a VFX gamble taken in the edit and became something directable on the day.
Digital resurrection
The practice predates the current AI wave — Peter Cushing’s Grand Moff Tarkin in Rogue One (2016) was conventional CGI over a performance by Guy Henry — but the toolset has changed. Alien: Romulus (2024) brought back Ian Holm’s Ash with the consent of his estate, and took criticism for it anyway. Consent from an estate settles the legal question and not the taste one.
Voice
Respeecher’s technology refined the Hungarian pronunciation in The Brutalist, a disclosure that became a minor awards-season scandal despite the performances being entirely Adrien Brody’s and Felicity Jones’s. The same company recreated Darth Vader’s voice for Star Wars with James Earl Jones’s blessing, and later his estate’s — which led directly to the sharpest test case yet.
The Fortnite charge
When an AI Vader voice appeared in Fortnite in 2025, SAG-AFTRA filed an unfair labour practice charge, arguing that the work should have been bargained for even though the estate had approved the voice. The estate’s permission and the union’s jurisdiction turned out to be two different questions — a distinction that will define the next several years of this argument.
![]()
Photo: Colin Smith, via Wikimedia Commons (CC BY-SA 2.0)
CGI vs AI: The Distinction That Keeps Getting Lost
It is worth being precise, because studios benefit from the vagueness.
CGI is constructed. Someone models the geometry, builds the rig, animates the performance, lights the shot and renders it. Every frame is the result of a decision a person can explain, revise and be credited for.
Generative AI is sampled. A model trained on an enormous quantity of existing material produces output that is statistically plausible. Direction happens through prompting and selection rather than construction, and the training material is where the ethical problem lives.
The practical differences show up on set. CGI is expensive, slow and precisely controllable. Generative work is cheap, fast and hard to control at the level of “move his left hand two inches”. Which is why the current production reality is hybrid: AI for iteration, look development, clean-up and volume; traditional VFX for anything that has to be exactly right.
Virtual Production Is Not AI — And Is Changing More
The technology that has changed how films are physically made this decade is virtual production: LED volumes running real-time game-engine environments behind the actors, so the background is photographed rather than composited.
There is no generative model in that pipeline. What there is, increasingly, is AI assistance around the edges — generating environment concepts, upscaling textures, filling in set extensions. For a performer, virtual production is unambiguously better than a green screen: you can see where you are. It gets swept into the AI conversation mainly because both arrived at once and both involve screens.
Deepfakes and the Consent Problem
The genuinely alarming end of this is not in cinemas. It is the trivial availability of tools that put a real face into footage they were never in.
The UK has criminalised the creation of sexually explicit deepfakes without consent, which addresses the worst category. Everything else — a dead actor endorsing a product, a working actor appearing in a film they never shot, a scam advert using a presenter’s face — sits in an awkward mix of passing off, data protection and defamation law that was not designed for any of this.
This is the strongest argument for the personality right the UK government floated in March: performers currently have to assemble a legal case from mismatched parts, while the harm is instant and the removal process is slow.
![]()
Photo: BalticServers.com, via Wikimedia Commons (CC BY-SA 3.0)
AI Restoration: The Use Nobody Objects To
For all the argument, there is one application almost everyone supports.
Machine learning is extremely good at the unglamorous work of film preservation: removing scratches and dirt, stabilising shrunken or warped frames, reducing grain without smearing detail, upscaling degraded elements and reconstructing damaged optical soundtracks. Work that took a specialist months of frame-by-frame cleaning now takes a fraction of that, which means smaller archives can afford to restore things at all.
The criticism here is aesthetic rather than ethical — over-smoothed faces, interpolated frames that make a 24fps film look like video, colours “corrected” away from what the cinematographer intended. Those are arguments about restraint, and they are worth having, but nobody’s consent is being violated by cleaning up a nitrate print.

Public domain, via Wikimedia Commons
What This Actually Means for Jobs
The honest answer is that the exposure is uneven, and the loudest fears are not aimed at the most likely losses.
A lead performance is not under threat in any near-term sense: audiences buy tickets to see specific people, and the cost of the star is not what makes a film expensive to produce. What is exposed is volume work — background artists, crowd replication, dubbing and localisation, pick-up lines, character voices in games, and the vast quantity of commercial and corporate content that never came near a union agreement.
Stunt performers sit in an unusual position: digital doubles can remove genuine physical risk, and also remove genuinely skilled jobs. The 2026 agreement addresses stunt work specifically, which tells you the union expects pressure there.
The other cost is less visible: the people whose past work trained the models. None of the consent architecture built since 2023 does anything about material scraped before it existed.
![]()
Photo: Phil Roeder, via Wikimedia Commons (CC BY 2.0)
What to Watch Next
- Whether the UK actually legislates a personality right, or whether March’s signal turns into another consultation.
- Whether any agency signs a synthetic performer. It has not happened yet, and the first one to do it will take the reputational hit for everyone.
- How the Fortnite charge resolves, because it sets the boundary between an estate’s permission and a union’s jurisdiction.
- Disclosure norms — whether credits and awards bodies settle on a standard for declaring generative tools, or leave audiences guessing.
- Streaming economics, where the volume of content being commissioned makes cheap synthetic filler most tempting.
The short version: the technology is real, most of it is being applied to real performers rather than replacing them, and the rules now covering that are stronger than they were in 2023 and still weaker than they need to be outside the union system. The first genuine AI film star is not here. The first thousand AI background artists probably are, and nobody announced them.
We cover this beat in AI & Future of Film, alongside our regular Movie News, Movie Reviews and OTT coverage. If you want to check what a specific film used, every title in our movie database carries cast, crew and production detail.
Sources
- SAG-AFTRA: Members Approve 2026 TV/Theatrical Contracts
- SAG-AFTRA: Artificial Intelligence resources
- The Hollywood Reporter: SAG-AFTRA files charge over AI Darth Vader in Fortnite
- The Hollywood Reporter: Metaphysic de-ageing in Here
- Deadline: Equity joins the backlash against Tilly Norwood
Frequently Asked Questions
What is an AI actor?
It is a loose term covering three different things: a digital replica of a real performer, a synthetic performer generated entirely by software with no human behind the face, and a character in an AI-generated video clip. The legal and union position on each is different, which is why the sloppy catch-all term causes so much confusion.
Is Tilly Norwood a real actress?
No. Tilly Norwood is a synthetic performer built by Xicoia, an AI talent studio spun out of Eline Van der Velden’s production company Particle6. There is no person being filmed. As of early 2026 she had not been signed by any talent agency, despite reported interest.
Has an AI actor ever been cast in a major film?
Not as a lead, no. What has happened repeatedly is the digital recreation or de-ageing of real actors — Tom Hanks and Robin Wright in Here, Peter Cushing’s Tarkin in Rogue One, Ian Holm’s Ash in Alien: Romulus. Those are replicas of real performances, not synthetic actors invented from nothing.
What did the 2026 SAG-AFTRA agreement change?
It tightened the rules on synthetic performers, which the union describes as ensuring synthetics remain the exception rather than the rule, and strengthened the consent and compensation framework around digital replicas won in 2023. Members ratified it with 91.42% in favour on a 19.25% turnout, and it runs from 1 July 2026 to 30 June 2030.
Can a studio use my face or voice without asking?
Not under a union agreement. Digital replica work requires informed consent to create the replica and separate consent for each use, plus payment. Outside union coverage — adverts, online content, foreign productions — the protection is far weaker, which is exactly the gap performers’ unions are lobbying to close.
What is the difference between CGI and AI in film?
CGI is built: an artist models, rigs, animates and lights something deliberately, frame by frame. Generative AI is sampled: a model trained on existing footage produces an output based on statistical patterns. The finished shot can look similar; the authorship, cost and control are completely different.
Does AI de-ageing count as an AI actor?
Usually not, in union terms. If the performer consented, was paid, and their own performance drives the shot, it is a digital replica of that performer rather than a synthetic one. The distinction matters for consent, credit and residuals.
Is virtual production the same as AI filmmaking?
No. Virtual production means shooting against LED volumes with real-time rendered environments, usually in a game engine. It predates the current AI wave and involves no generative model. The two are increasingly used on the same stage, which is why they get conflated.
What is a deepfake, and is it legal in the UK?
A deepfake is a synthetic video or audio clip that puts a real person’s face or voice into something they never did. Making sexually explicit deepfakes without consent is a criminal offence in the UK. Other uses sit in a murkier area of passing off, data protection and, increasingly, the personality right the government has said it is considering.
Are AI actors cheaper than human ones?
For a headline performance, no — the expensive parts of a film are still development, crew, locations and post. Where the sums do change is in volume work: crowd replacement, background extras, dubbing, pick-up lines and social content. That is where the job losses are most plausible.
Which films have used AI voice technology?
The most discussed case is The Brutalist, where Respeecher was used to refine the Hungarian pronunciation of dialogue spoken by Adrien Brody and Felicity Jones. Star Wars has used the same company’s technology to recreate Darth Vader’s voice, with James Earl Jones’s consent and later his estate’s.
Why did SAG-AFTRA take action over Darth Vader in Fortnite?
In 2025 the union filed an unfair labour practice charge over the AI-generated Vader voice added to Fortnite, arguing that the work should have been bargained for even though the estate had approved the use of the voice. It became the clearest test of who has to be consulted when a dead performer’s voice returns.
Can an AI-generated performance be copyrighted?
Output generated purely by a machine, with no meaningful human authorship, generally cannot be copyrighted in the US, and the UK position on computer-generated works is under active review. In practice studios avoid the question by keeping human authorship clearly in the loop.
What did the UK government decide about AI and copyright?
In March 2026 it stepped back from the broad opt-out exception it had previously favoured, saying it would not reform copyright law until it was confident the changes met its objectives, and floated a new personality right covering unauthorised use of a performer’s voice and likeness.
What is Equity’s position on AI?
Equity has campaigned hard against unlicensed use of members’ voices and images, called the proposed opt-out exception “legalising theft”, and welcomed the government’s change of course in 2026. It also joined the backlash against Tilly Norwood.
Will AI replace background actors first?
That is the widely shared fear, and it is the most economically exposed category, along with voice work for dubbing and games. Union agreements now require consent and payment for scanning background performers, which slows it down but does not remove the incentive.
Can AI restore old films?
Yes, and this is the least contested use of the technology. Machine learning is very good at removing scratches, stabilising warped frames, upscaling, reducing grain and reconstructing damaged audio. Purists’ objections are about over-processing, not consent.
How can I tell if what I am watching used AI?
Often you cannot, which is part of the problem. Credits are improving slowly, and some festivals and awards bodies now ask for disclosure of generative tools. Union agreements increasingly require notice when a synthetic performer is used.
Does an AI actor need to be paid?
The synthetic performer itself, obviously not — that is the point of it commercially. Where payment questions arise is the training data: the performers whose work taught the model, who under current law were mostly neither asked nor paid.
Where can I follow this story on Filmoria?
Our AI & Future of Film section covers AI-generated films, AI actors, virtual production, deepfakes, restoration and the copyright fights around them, alongside the regular Movie News and OTT coverage.
