Deepfakes in Hollywood: How AI Faces and Voices Are Changing Film

A Hollywood deepfake usually means one of two things. On set, it's AI-assisted face or voice work done with consent and blended with traditional VFX, like the younger Harrison Ford in Indiana Jones and the Dial of Destiny or the AI-built Val Kilmer in As Deep as the Grave. Online, it's an unauthorised fake of a star's face or voice, like the Seedance 2.0 clips of Tom Cruise fighting Brad Pitt that Hollywood condemned in February 2026. The first is a craft tool. The second is the problem, and in 2026 it's become much harder to catch by eye or by ear.

Cinema has always been about pulling you into a convincing illusion. In the last few years, that illusion has started to shift in a more literal direction, as artificial intelligence becomes part of how faces, voices, and performances are constructed on screen.

The term “deepfake” gets used loosely, but what Hollywood is really adopting is a growing set of AI-assisted tools for digital likeness, face replacement, voice cloning and performance reconstruction. These techniques are changing how films are made, raising new ethical questions, and quietly redrawing the boundary between what is recorded and what is generated.

This post covers how the technology works in real productions, which films have used it, why Seedance became a flashpoint, how far the 2026 models and voice clones have come, whether anyone can still spot a fake, where the law stands, and what all of this means for storytelling.

What is a deepfake in film?

At a technical level, deepfakes are a subset of synthetic media created using machine-learning models trained on large volumes of visual and audio data. Early systems relied heavily on generative adversarial networks, but modern pipelines increasingly use diffusion-based and transformer-driven models for face synthesis, voice cloning, and lip-sync. The newest models generate picture and sound together, so a fake can now arrive as a complete clip rather than a face pasted onto someone else.

In practical terms, the technology allows a person’s likeness or voice to be reconstructed or altered in ways that can look convincingly real. In entertainment, these systems rarely operate alone. They are layered into existing visual effects workflows alongside motion capture, 3D modelling, compositing, and manual animation.

That distinction matters. What audiences see on screen is not a raw deepfake in the internet sense, but a hybrid of AI and traditional VFX craft.

Which Hollywood films have used deepfakes or AI faces?

Recent high-profile films are often described as “deepfake movies,” but the reality is more nuanced. Here's what was actually done in the films people ask about most.

Sources: fxguide on The Irishman, Harbor on the making of Here, AP on As Deep as the Grave, and the ILM breakdown below.

Indiana Jones and the Dial of Destiny (2023) used machine-learning-assisted de-aging and face replacement tools combined with traditional VFX techniques to create a younger version of Harrison Ford. Lucasfilm described the process as an internal system trained on decades of archival footage rather than a consumer-style deepfake tool.

Here (2024) goes further. Its VFX supervisor, Kevin Baillie, says about 53 minutes of the final film needed full digital face replacement of the four leads, and that traditional CGI would have been too slow and expensive, so the team relied heavily on machine learning and AI-based techniques (Harbor).

As Deep as the Grave (2026) is the one that moves the line. Val Kilmer was cast as Father Fintan but was too ill to shoot his scenes, and he died in April 2025. The filmmakers used generative AI to complete the role, and his estate approved the project, was paid, and supplied material for the recreation, according to the Associated Press. Director Coerte Voorhees told the New York Times the likeness was built with a mix of publicly available generative AI tools (via the San Juan Daily Star). Kilmer had been here before. After throat cancer took his voice, software rebuilt it from his own recordings, and that voice was used in Top Gun: Maverick (AP).

Rogue One: A Star Wars Story (2016) recreated Peter Cushing’s likeness using CGI, performance capture, and extensive manual animation. No deepfake system was involved, but it remains a foundational reference point for modern digital-likeness debates.

Fast X (2023) included Paul Walker through a blend of archive footage, body doubles, and CGI face work. Again, this was not deepfake technology, but it sits on the same ethical and creative fault line.

What has changed since those films is not the concept, but the accessibility and speed of face-synthesis tools. Seven years ago, de-ageing three leads took ILM a bespoke camera rig. In 2026, a posthumous lead performance was built with tools anyone can subscribe to, even if finished film work still needs careful supervision.

The rules on all of this vary from country to country, and the technology is moving so fast that governance is still catching up. So I treat consent as my job, whatever the law currently allows. I set out how in my guide to getting consent for a digital replica or AI voice.

What is the problem with deepfakes in Hollywood?

This is where the conversation becomes more serious than visual novelty.

Consent and control

The 2023 SAG-AFTRA strike pushed digital likeness rights into the centre of contract negotiations. Since then, many new agreements include explicit clauses covering how digital replicas and likeness can be licensed and reused. This is now standard practice rather than a niche concern, and recent reporting on actors fighting back as AI deepfakes become scarily accurate shows how those concerns are now extending beyond contract language into active efforts to protect consent, attribution, and likeness control.

If you're a producer, I've set out the practical side in securing likeness and voice permissions before you shoot.

This shift also ties into a wider industry reckoning about ownership and leverage, something I explored in my analysis of how actors are increasingly being treated as long-term intellectual property rather than just performers.

Posthumous performances

Digitally recreating deceased performers remains legally complex and culturally divisive. As Deep as the Grave shows what the careful version looks like: the estate agreed, was paid, and helped supply the material. Even then, it raises unresolved questions about who owns a legacy and where artistic homage ends and commercial exploitation begins. California now requires an estate's consent before a digital replica of a dead performer is used commercially (see the law section below).

Misinformation and reputational harm

Outside cinema, synthetic video and audio have already been used in scams and political misinformation. That broader context is one reason lawmakers in the US, the EU, the UK and Denmark have begun drafting or passing laws targeting non-consensual or deceptive synthetic media.

For film studios, this creates a reputational risk. The deeper harm is not only the fake, it’s the shift into viewers watching with their guard up, because once suspicion becomes the default the whole viewing experience changes. That’s why proof of origin and clear responsibility matter. The same tools that make production more flexible also increase the need for transparency and consent frameworks that audiences can trust.

Seedance and the 2026 backlash

The moment this stopped being a niche debate was February 2026. ByteDance, the company behind TikTok, launched its Seedance 2.0 video model on 12 February, and within days clips of Tom Cruise fighting Brad Pitt, and riffs on Stranger Things and The Avengers, were all over social media. The Human Artistry Campaign, a coalition that includes SAG-AFTRA and the Directors Guild of America, called it destructive, and the Motion Picture Association told ByteDance to “immediately cease its infringing activity” (Deadline). Disney sent ByteDance a cease-and-desist letter on 13 February (Axios).

The current version is Seedance 2.5, which ByteDance launched on 31 July 2026. It makes clips of up to 30 seconds, with sound, in one go, and can take up to 30 reference images, 10 video clips and 10 audio clips at once (ByteDance Seed). In August the MPA signed its first agreement with an AI company, a framework with ByteDance to protect film and TV intellectual property across Seedance and its image tool Seedream (Variety). Testing reported by Deadline found Seedance 2.5 blocked around 80% of attempts to recreate copyrighted material, but testers still produced a faithful copy of Oscar from Shark Tale without naming him (Deadline).

So the guardrails are better, but they're mainly about studios' characters and footage, and they aren't watertight. For anyone making video, the lesson isn't that these tools are off-limits. It's that the toolmaker's filters are not your consent process. That part is still yours.

How good the 2026 AI video models have got

Seedance isn't the only model that's moved fast. Not long ago, most AI video was a few silent seconds with a giveaway somewhere in the frame. The leading 2026 models make picture and sound together, hold a character's face from shot to shot, and run long enough to carry a scene.

Kuaishou launched Kling 4.0 on 15 October 2026, with steadier motion and smoother tracking and orbit shots (Bernama) and single clips of up to 30 seconds (fal). Google's Gemini Omni Flash, released in May, goes further in a different direction. You can shoot your own footage and ask it to change what happens, add new characters, or switch the setting, and it can build an avatar that looks and sounds like you from your own voice. Google says every Omni video carries its invisible SynthID watermark, and that it's still testing the ability to change speech in existing video before releasing it (Engadget).

That editing point matters more than any spec sheet. A deepfake used to mean one person's face on another person's body. Now it can be real footage, re-directed after the fact, with nothing obviously wrong in the frame.

Not everyone stayed in the race. OpenAI announced in March that it was shutting Sora down, closed the app on 26 April 2026 and switched off the API on 24 September (The Decoder). The model that started the public conversation about AI video is gone, and the tools that replaced it are better.

Voice cloning: the deepfake you hear

Faces get the headlines, but voice is moving faster. A voice is easier to lift out of a performance than a whole face and body, and it takes far less data to copy convincingly.

Licensed voice clones are now a business. Michael Caine put his voice on ElevenLabs' Iconic Marketplace, where brands and producers can ask to license it, and Matthew McConaughey uses the same company to put out a Spanish edition of his newsletter in his own voice (ElevenLabs).

The other side is messier. The Los Angeles Times reported in August 2026 that voice actors are split, with established names earning well from licensed clones while working performers watch jobs disappear, and that the National Association of Voice Actors lists more than 100 actors who've found clones of their voices online (LA Times). In the UK, more than 80 performers, including Hugh Bonneville, Nicola Coughlan and Matt Lucas, backed the Save Our Voices Now campaign in August, calling for a law that protects people's voices. The campaign says 28% of UK adults report being targeted by a voice-cloning scam (Express Tribune).

There's a quieter cost too. A 2026 study of nearly 1,800 listeners found people were no worse at catching fake speech than they had been in 2021, but noticeably worse at trusting real speech, marking genuine recordings as fake far more often (arXiv). That's the damage that doesn't show up in a scam report. Once any voice can be cloned, every voice is under suspicion, including the real ones.

I've looked at what this means for creative work in AI voiceovers, threat or new frontier.

Can you still spot a deepfake?

Mostly, no. In DF26, a benchmark published in September 2026 using video from seven of the year's generators, people correctly picked out the fakes about 53% of the time, barely better than a coin toss, and several leading automated detectors scored close to chance on the newest material (DF26). In the US arm of a Veriff and Kantar study, around half of people said they could reliably spot manipulated media, but scores sat close to guessing whether they were confident or not (Veriff).

So the effort is shifting from catching fakes to proving what's real. In May 2026 OpenAI joined the steering committee of C2PA, the standard behind Content Credentials, and said it would add Google's SynthID watermark to its images, while Google said it's bringing Content Credentials and SynthID checks to Search and Chrome (C2PA Viewer). Content Credentials work like a signed history attached to a file. SynthID is a watermark hidden in the picture or sound itself.

Neither is a fix on its own. Content Credentials can be lost when a file is re-exported or posted somewhere that doesn't read them, and SynthID only marks content from tools that use it (Trust Insights). Missing credentials don't prove a clip is fake. What these tools do give you is a way to show your own work is what you say it is, and that's the approach I take. How I label AI in my own work is set out in my AI policy.

Are deepfakes legal? Where the law stands in 2026

Short answer: using a real person's face or voice without consent is increasingly risky, but the rules depend on where you are. Correct as of 30 October 2026. This is a plain summary, not legal advice. Rules differ by country, so check local law or take advice for a real project.

United States

SAG-AFTRA members ratified a new TV and theatrical agreement in 2026 by 91.42%. It runs from 1 July 2026 to 30 June 2030 (SHOOTonline). Under it, producers can't use fully synthetic performers without first bargaining with the union, they agree to a principle “strongly favoring human performances”, and they don't intend to put a synthetic in a human role unless it brings “significant additional value” (SAG-AFTRA).

California already requires contracts to spell out any use of a performer's digital replica (AB 2602), and bans commercial use of a dead performer's replica without the estate's consent (AB 1836). Both were signed in September 2024 (Governor of California).

At federal level, the NO FAKES Act would give everyone the right to authorise or refuse digital replicas of their voice and likeness, with exceptions for news, documentary, parody and biographical work, and a right after death of up to 70 years. It cleared the Senate Judiciary Committee on 18 June 2026 (Deadline). On 30 September, Senator Marsha Blackburn tried to pass it by unanimous consent and Senator Ted Cruz objected, saying its protections for satire and political speech weren't strong enough. That leaves it waiting for the short session after the November midterms (Radio Ink), or a fresh start in the next Congress (Beyond Copying). It isn't law yet.

European Union

Since 2 August 2026, Article 50 of the EU AI Act has required deepfakes to be disclosed. For clearly artistic, creative or fictional work, the disclosure only has to be appropriate and not spoil the work. The European Commission also says AI used for background scenes, special effects and standard post-production is not likely to make content falsely appear authentic (European Commission).

United Kingdom

Since February 2026, creating or requesting an intimate deepfake of an adult without consent has been a criminal offence in England and Wales (Data (Use and Access) Act 2025, section 138). Beyond that, the UK has no general right over your own face or voice. Media lawyer Dr Mathilde Pavis calls the law “unfit for purpose” (BBC). In March 2026 the government promised a consultation on digital replicas, with a new personality right among the options it will look at (Hogan Lovells). Equity is calling for personality rights, and says its talks with Pact on AI protections for UK film and TV performers are at an advanced stage (Equity).

Elsewhere

Denmark introduced a bill on 8 October 2026 to ban sharing realistic deepfakes of people's faces and voices without consent. The minister hopes it will take effect at the start of 2027 (Euronews).

Will deepfakes replace actors?

In practice, AI tools are being absorbed into filmmaking the same way digital compositing and CGI once were. They remove friction from certain tasks, but they do not replace the creative judgement that shapes a performance or a story.

Where they already make practical sense:

• correcting continuity issues without reshoots

• reducing the need for dangerous or physically demanding stunts

• extending limited archival footage for flashbacks or historical scenes

Where they still fall short:

• conveying subtle emotional shifts

• improvisational nuance

• the unpredictable energy of live performance

The result is not a replacement of actors or directors, but a quiet rebalancing of what is captured physically and what is constructed digitally. The 2026 SAG-AFTRA deal points the same way, with producers committing to a principle that favours human performance.

The test case is Tilly Norwood, the AI "actor" created by London studio Particle6, which unions including SAG-AFTRA and Equity have condemned. In July 2026 Particle6 announced she'll lead a feature, Misaligned, made by human writers, directors and editors working alongside AI specialists (Variety). It's still in development. Whether an audience will connect with a lead who has never lived anything is the question that film will have to answer.

This shift is less about spectacle and more about workflow. It changes how scenes are finished, how long fixes take, and how much of a performance can be shaped after the camera stops rolling.

There is also a quieter shift underway. What begins as AI acting as a collaborator on technical problems will, over time, edge toward creative influence. Not in the sense of replacing writers or actors, but in subtly shaping what performances are possible, what edits are practical, and what visual ideas survive the cost and time constraints of production.

The more invisible that influence becomes, the more it will matter. When audiences stop noticing whether a tear was captured on set or generated in post, the question is no longer about realism. It is about whether storytelling becomes more expressive and daring, or more uniform and risk-averse.

What has changed since the 2023 strike

Three structural shifts now define how synthetic media is actually being absorbed into film and media production.

First, digital likeness rights have moved from abstract principle into routine contract practice. Studios increasingly treat consent, scanning, reuse, and licensing terms as core legal infrastructure rather than optional add-ons negotiated at the margins.

Second, detection has fallen behind. Research groups and governments have spent years funding forensic tools, but the 2026 generation of video and voice models beats most of them. The focus has moved to provenance instead: watermarks and signed records that show where a file came from and what was done to it.

Third, the barrier to entry has collapsed. What once required a major visual effects house can now be rough-tested by small teams using off-the-shelf tools, and a face, a voice and a soundtrack can come out of a single generation. Seedance showed how fast that can spill into public view. That makes governance, consent, and transparency more important, not less.

What Hollywood deepfakes mean for brands and businesses

Hollywood is effectively stress-testing technologies that will appear in marketing, training, and communications.

Possible uses already being explored in industry:

• multilingual video localisation using cloned voices

• historical reconstructions for education

• digital presenters for internal corporate media

The risk is not technical failure. It is erosion of trust. When synthetic video becomes visually flawless, the audience’s first question stops being “is this impressive” and becomes “is this real”. If you're cloning a voice, even your own chief executive's, get the consent in writing and say so when you publish.

If you're planning anything like this, start with a working standard for using AI ethically in video. And if you'd like help making it, here's how I approach AI video production.

What deepfakes mean for the future of cinema

Deepfakes are not a gimmick. They are part of a deeper shift in how reality itself is represented and mediated. For filmmakers, the challenge is not technical. It is ethical and narrative.

The real risk is not that AI will replace creativity. It is that it will make creative choices feel cheaper, safer, and more homogenised if used without intention. Tools are neutral. Direction is not.

The directors and studios who use AI to take bigger creative risks will define the next era of cinema. Those who use it only to reduce cost and friction will produce work that looks impressive and feels empty.

Deepfakes will not kill cinema.
Bland reliance on them might.

Questions people ask about Hollywood deepfakes

What is a Hollywood deepfake?

It usually means one of two things. On set, it's AI-assisted face or voice work done with consent and blended with traditional VFX, as with the younger Harrison Ford in Indiana Jones and the Dial of Destiny. Online, it's an unauthorised fake of a star's face or voice. The first is a craft tool. The second is the problem.

Which films have used deepfake or AI face technology?

Here (2024) used machine learning and AI-based face replacement for about 53 minutes of the film. As Deep as the Grave (2026) used generative AI, with his estate's consent, to recreate Val Kilmer after his death. The Irishman (2019) and Indiana Jones and the Dial of Destiny (2023) used AI-assisted tools inside traditional VFX to de-age their leads. Rogue One (2016) and Fast X (2023) are often called deepfakes, but used CGI, doubles and archive footage.

Can you still tell if a video is a deepfake?

Not reliably by eye. In a September 2026 study using that year's video models, people spotted the fakes only about 53% of the time, and most automated detectors struggled too. A better check is provenance: look for Content Credentials or a SynthID watermark, and treat missing credentials as unknown rather than proof of a fake.

Is it legal to deepfake an actor?

Not without consent in most professional settings. In the US, the 2026 SAG-AFTRA deal and California law require consent and clear contract terms for digital replicas, although there's no federal law yet. In the EU, deepfakes must be disclosed. The UK has no general right over your own face or voice yet, although intimate deepfakes made without consent are now a crime in England and Wales.

Is AI voice cloning legal?

Cloning your own voice is fine. Cloning someone else's without consent is risky almost everywhere. In the US, California law and the 2026 SAG-AFTRA deal require consent and clear terms for performers' digital replicas, and the federal NO FAKES Act is still pending. The UK has no specific right over your voice yet, which is what the Save Our Voices Now campaign wants to change.

What is the NO FAKES Act, and is it law yet?

It's a proposed US federal law that would give people the right to authorise or refuse digital replicas of their voice and likeness, with exceptions for news, documentary, parody and biographical work. It cleared the Senate Judiciary Committee on 18 June 2026, but on 30 September Senator Ted Cruz blocked a bid to pass it quickly, citing free-speech concerns. It isn't law yet.

Do films have to label deepfakes?

In the EU, yes, in a light-touch way. Since 2 August 2026, the AI Act requires deepfakes to be disclosed, but for clearly artistic or fictional work the disclosure only has to be appropriate and not spoil the film. The European Commission says AI used for background scenes and standard effects is not likely to count as a misleading deepfake.