How Trey Parker and Matt Stone Built Deep Voodoo Into an AI Company Hollywood Can Actually Trust

A brick-walled studio in Venice, and not much drama until the cameras stop

In a neatly lit brick-walled space in Venice, actors walk in, pose for photos, and sit for video capture. It is the sort of efficient, mildly anonymous production setup Los Angeles has perfected over decades. The process looks ordinary enough. Then the footage leaves the room, and things get weird in the highly organized way only Hollywood can manage.

At Deep Voodoo, the boutique company founded by South Park creators Trey Parker and Matt Stone, those images are converted into data and sent to AI specialists scattered around the world, including in Eastern Europe, Argentina, and Vancouver. With the help of computing power from an undisclosed data center, the material is turned into synthetic performances, de-aging effects, deepfakes, and other face-swapping visuals for entertainment projects.

Stone, 54, says that is the point. He is unimpressed by the usual AI sales pitch about replacing everyday tasks. “I find that a lot of discussions about AI become tiresome. You know ‘put your taxes in and it can do them,’” he says. “And it’s like, ‘cool, but a human can do your taxes.’ What we’re trying to do is something no amount of humans can do.”

That ambition has already shown up in more places than many viewers realize.

The work you may have already seen

Deep Voodoo has quietly contributed to a string of high-profile projects, some obvious in hindsight and some not at all. Among them:

• The Kendrick Lamar music video from a few years ago in which Lamar’s face shifts into O.J. Simpson, Will Smith, and Jussie Smollett
• The Bill Clinton food-counter gag in Ted earlier this month
• The Ben Affleck and company 1990s nostalgia spot for Dunkin’ Donuts that ran during the Super Bowl last month
• The Donald Trump full-frontal deepfake in the South Park season 27 premiere last summer

With generative AI moving closer to a permanent place in Hollywood production, Deep Voodoo’s work is likely to become more familiar, whether audiences know it or not. If a studio needs something that can morph, shift, or swap faces without collapsing into obvious fakery, Parker and Stone are increasingly the people making the call, or at least answering it.

And, oddly enough, they appear to be trying to do it without setting the entire internet on fire. A rare commitment in this sector.

It started because a Trump deepfake movie did not have the right tech

Deep Voodoo was not originally meant to become a company. It emerged because Parker and Stone ran into a problem while developing a Donald Trump deepfake film during the late years of the first Trump administration.

The concept was to place Trump’s face onto another actor’s body, then let the character deteriorate further and further, eventually losing both his sanity and, later, his clothes. But the pair could not find a studio that could deliver the level of effects they wanted.

“A couple of effects houses in LA just kind of gave us the runaround,” Stone says. “This has happened before in our career, where we go, ‘okay, well, we’ve got to go figure it out ourselves.’” So they did what stubborn creators do when the industry stalls them, they found AI specialists online and assembled their own team.

The movie never happened. Covid killed it before it could get off the ground. But the team did not disappear.

Instead, it became the basis for Sassy Justice, a web series that spoofed public figures. A 14-minute episode featuring a deepfaked Trump spread widely online. By 2021 standards, the visuals and audio were rough by current expectations, but they were still remarkable at the time, and Parker and Stone later reused some of the material for the South Park season opener that July.

By late 2022, Deep Voodoo had become a real company, and not just a clever side project. It had raised $20 million, including money from a venture effort tied to CAA, at a moment when Hollywood had not yet fully started arguing with itself about AI in public.

Why the company sounds less like Silicon Valley and more like a compliance memo

Deep Voodoo seems designed to avoid the usual startup theater. Its leadership is unusually calm for a company working in one of the most controversial corners of AI.

Jennifer Howell, an animation veteran who previously produced South Park and worked at a long list of studios, is chief content officer. Afshin Beyzaee, the CEO, is an unflashy lawyer who spent years as chief counsel at Parker and Stone’s Park County production company.

Neither sounds interested in tech bro mythology. Beyzaee is more likely to talk about rights and permissions than disruption, which is refreshing in the same way a seat belt is refreshing. He says flatly that it is “very inappropriate to be taking and making use of someone’s likeness without their permission.”

That attitude is central to how the company operates. Deep Voodoo says it will not work with studios unless actors or estates have authorized use of their likenesses. The only notable exception raised in the story is the Trump deepfake used last summer, which did not have White House permission. Executives said they relied on fair-use images of the former president, who has never exactly been hard to find in public life.

“You’ve got a situation that some are paying to use or license IP and some are not and then people say ‘why should I pay for it?’” Beyzaee says. “To us, it’s this is about making sure that we’re providing this service, providing this technology, in a way that respects the laws and the protections and the rights that people have.”

The company also says it can walk away from jobs if the permissions are not sufficient.

The method: licensed footage, a lot of cameras, and a month of work

Deep Voodoo does not scrape the web for images and then train a model on whatever it finds. That is not the company’s pitch, and it is not how it says it works.

Instead, it uses licensed material, either captured in its Venice studio or provided by the production company. In the studio, an actor is filmed with nine cameras and asked a sequence of simple questions designed to produce different facial reactions. That data is then used to build a custom model for the specific project.

It is slow, expensive, and not remotely the kind of process a scale-obsessed Silicon Valley firm would celebrate at a conference. A single project can take up to a month and involve about 300,000 images. But the result is tailored to the production and, crucially, legally cleared.

“It’s important that this isn’t us just going and scouring the Internet for materials and building it into our models,” Beyzaee says.

That difference matters to the company, and it also matters to the productions it serves. When Deep Voodoo de-aged, and then re-aged, Billy Joel for the 2024 video for “Turn The Lights Back On,” the effect worked because the production had a system built specifically for that job instead of a generic model stitched together from random online content. The shifting between decades was meant to feel seamless, and in that case, it did.

“Our goal is to make beautiful, cinematic film and television that never pulls the viewer out because the effect doesn’t look right,” Howell says. The company, she adds, was started by artists who are extremely picky. Hollywood, naturally, has never had any shortage of those.

The creative argument, and the AI argument, are not the same thing

Deep Voodoo’s team knows the entertainment industry is skeptical of AI, especially among actors and writers. But they argue that a lot of the backlash is aimed at prompt-based tools that try to create content without a performer at the center.

Stone draws a line between that kind of synthetic output and what Deep Voodoo does. “We’re not doing anything like that. We’re not typing in a prompt. It’s all capturing actors doing what they do,” he says.

He compares the process to puppetry. The tool can make the puppet, but the performance is what gives it life. “I mean, the magical part of the production is the puppeteer, right? The puppet is one thing, and the tools can create a great puppet. But the magic is the performer. Without that is just becomes wallpaper.”

That is also what makes the company such an awkward fit for the broader AI debate. It is a firm built around synthetic images, but it insists on consent, performance, and authorship. In other words, it is trying to be the good version of something most people are still trying to understand.

What comes next: more de-aging, more performance transfer, and probably more fights about it

Stone thinks the technology is still in the early stages of what it can do.

He expects it to power horror movies, comedies, and political shows built around exaggerated, current events-driven deepfake satire. In his view, the format could eventually support something like an SNL-style weekly or biweekly schedule, with grotesque hybrids of real figures rather than straight impersonations.

“Somebody’s going to make a scary fucking horror movie using this technology. Somebody’s going to make a really fucking funny comedy using this. Like really funny shit that couldn’t be made,” he says. “What’s going to happen soon, what we’re capable of, is someone’s going to do a political show.”

De-aging remains one of Deep Voodoo’s strongest use cases, but Howell and Stone also point to “performance transfer,” a technique that lets actors perform in a controlled setting, often in street clothes on a stage, with limited location shooting. Their performance is then transferred into a scene that looks as if it took place somewhere else entirely, such as the streets of Paris or a confrontation in Beijing. Think of it as a very elaborate version of re-recorded dialogue, except the body is invited too.

The potential cost savings are obvious. So is the threat to location crews and to the cities and countries that spend time and money luring productions. Stone does not dismiss that. He acknowledges the downside, especially for the physical production ecosystem.

He also argues the change is unavoidable.

“We’re already all watched stuff on TV that has utilized machine learning,” he says. “It’s happening, and it is going to change the industry.”

As for South Park, the answer is probably yes, eventually.

The show now runs on a two-week cycle, a pace Stone says has more to do with age than technology. Still, he thinks the tools may help. “Maybe we go home earlier, maybe we’ve got more options,” he says. “And it means maybe the show’s better.”

For a company built by people famous for pushing every possible boundary, Deep Voodoo’s most surprising move may be insisting on rules. That, and trying to make the future of AI look less like a free-for-all and more like a production schedule.