The Tom Cruise deepfake that set off 'terror' in the heart of Washington DC

It only became clear what Chris Ume had unleashed in the days after his TikTok Tom Cruise videos went viral in February this year. Then fear started to set in. 

Suddenly it felt like the whole world was watching. In a few days, one of the videos, seemingly showing the Hollywood superstar teeing off at the golf course, already had 2.5 million views. 

Then it kept climbing to 10 million and beyond. TikTok Tom was becoming a global phenomenon.

"In the beginning it was really scary ... we never expected it," the Belgian-born graphics artist tells Foreign Correspondent from his current home in Bangkok.

"We started talking about it like, what should we do?" 

Ume's collaborator fired off an email to Cruise's management, asking if they should take the videos down and offering to hand over the @deeptomcruise TikTok account.

It wasn't that the videos were offensive, or scandalous, or even damaging to Tom Cruise. The actor's management never even wrote back. It's that they were good. Scary good.

On the other side of the world, in Washington DC, TikTok Tom was soon setting off alarm bells in national security and intelligence circles. One insider described the response as nothing short of "terror".

A series of clips depicting one of the most recognised faces in the world, created by one guy at home with a computer, had fooled virtually every piece of publicly available deepfake-detection technology.

Deepfakes are Artificial Intelligence-generated videos which can make people appear to do and say things they never did.

Ever since the first low-quality deepfakes emerged five years ago, the technology has been viewed as an emerging threat to national security, liberal democracy and even truth itself.

In 2021, that threat has come much more sharply into focus, as examples such as the Tom Cruise video show the technology is now ready to produce lifelike fakes.

Few would have predicted that the harbingers of a future some speak of in apocalyptic terms would be a graphics wiz from Belgium and a Tom Cruise impersonator.

The making of Tom Cruise

It could almost be Tom Cruise. There's the same action-hero gait, the familiar gestures and the megawatt smile. There's even an unmistakable note of the Cruise rasp in his voice. 

But the man in the video is not Tom Cruise, it's Miles Fisher, an actor who created a spoof campaign clip called "Run Tom Run" in the lead-up to the 2020 US presidential election.

"It was a funny video where Tom Cruise was running for president," Ume says. "But it was just [Fisher], it was no deepfake."

When Fisher started planning the sequel, he turned to Ume for help. Fisher wanted Tom's "concession" speech to use deepfake technology for the big reveal, where Tom Cruise would seem to emerge dripping wet from a swimming pool, laughing manically and draped in an American flag.

The resulting video, uploaded to social media, got the reaction the pair were after.

"People were amazed, like, 'How did he do this?'" Ume recalls. 

Their partnership would soon go on to produce some of the most convincing deepfakes ever made.

Ume first got interested in deepfakes in late 2018 when he saw a news report about how the technology was being used for malicious ends. Where others were exploring the darker side of deepfakes, Ume saw the creative potential. 

"I did a few months of research, how to do it, and a half-year later I had my first deepfake," he says. "Then it evolved rather quickly. I just started creating crazy things."

Those "crazy things" soon included a long-haired, preening Snoop Dogg. Ume is now working full time on deepfakes.

It's estimated 93 per cent of deepfakes currently produced are pornography, but the technology is rapidly emerging from the dark corners of the web into the mainstream.

Graphics and AI experts like Ume are expanding the possibilities of deepfakes into the entertainment and commercial realms.

A deepfaked David Beckham has fronted a malaria eradication campaign, delivering the message in nine languages. 

Apps and online tools have emerged offering users everything from a personalised video message from football legend Lionel Messi, to "deep nostalgia", where photos of dead loved ones can be uploaded to recreate their likeness in a moving video.

The number of professionally made deepfakes is now doubling every six months. Some experts are predicting up to 90 per cent of online videos could be synthetically generated — deepfakes — by 2030.

Hollywood is also taking notice.

Last year, after leaving Belgium, Ume got a break working alongside other handpicked deepfake wizards on a new TV series by South Park creators Matt Stone and Trey Parker. 

Called Sassy Justice, the show is a spoof local news report anchored by Fred Sassy, a deepfake who looks like Donald Trump behind a cravat and a wig of wiry grey curls.

Some of Ume's early work gained notoriety, like a deepfake of Kit Harrington, the actor who plays Jon Snow in Game of Thrones, apologising for the show's final season, and a digital restoration of a performance by renowned Belgian singer Jacques Brel. 

Both used Ume's' visual effects wizardry to manipulate pre-existing footage. But the near-flawless deepfake Tom Cruise videos were made from scratch, the product of an art-meets-science approach.

Fisher, who Ume rates as one of the "world's best" Tom Cruise impersonators, acted out scenes for the camera, channelling the Top Gun star's voice and mannerisms. It was then up to Ume to develop the AI model to replace his face with Tom Cruise's.

"Miles looks a lot like Tom Cruise, he has a lot of similarities," says Ume. "That makes my work easier in a way. It saves me a lot of time."

The key to building a deepfake is what's called a "generative adversarial network". It requires two neural networks — computer algorithms that mimic the way our brains work — which learn from each other and adapt. The two neural networks are then pitted against each other to build a perfect deepfake.

"For a deepfake, you start with the source data — I mean pictures and videos of the character you want to deepfake. In my case, Tom Cruise."

Ume collected over 6,000 images of the Hollywood actor taken from different angles, with different facial expressions, to train the algorithm. It still took over two months to build the computer model, then dozens more hours ironing out the glitches frame by frame using digital graphics software.

Then Ume and Fisher released their creations into the wilds of the internet. The reaction was completely unexpected, as online viewers debated whether the real Tom Cruise had in fact joined TikTok.

"Because we never exposed ourselves in the beginning, people didn't know who created these videos," Ume says. "You had a lot of articles talking about the end of the world … and, 'This technology is getting out of hand.'"

To quell some of the hysteria, Ume went on a media blitz, revealing how he made the videos and reassuring the world his intentions were only to entertain. 

Since then, Ume has been talking to governments too, advising them on the rapid evolution of deepfake capability.

"A lot of governments still don't know what deepfakes are and I'm here to explain what it is, which things are possible," he says.

"And governments realise this could be used as a weapon."

Weaponising deepfakes

It's a possibility that keeps former CIA officer and disinformation specialist Matt Ferraro up at night.

When deepfake Tom started doing the rounds in Washington DC's national security and intelligence community, some feared a dystopian future had taken one giant leap closer. Matt was one of them.

"I think that 'terror' is probably not too strong a word," says Ferraro, who worked for America's top spymaster, the director of National Intelligence, during the Bush and Obama administrations. 

"It's because they realise how dangerous [deepfakes] are. It does seem like it's really going to be a fundamental challenge to the information environment."

When Ferraro started writing about deepfakes two years ago, most were easily detected by the naked eye. But that has changed dramatically in the past 18 months, he says. Today's deepfakes have become "radically good" and now pose an imminent risk to the political system.

Early deepfakes, such as filmmaker and actor Jordan Peele's 2018 Obama video — in which a synthesised likeness of the former president calls his successor a "total and complete dipshit" and then signs off with the straight-faced salutation "stay woke, bitches" — highlighted the potential for deepfakes to sow political chaos. But the visual sophistication was lacking, Ferraro says.

"It was mostly old video of Barack Obama with only his lips moving. You could tell that was fake. Now with the Tom Cruise deepfakes, it's getting harder and harder".

The spectre of a deepfaked political leader has already arisen in Africa, where a video of Gabon's reclusive president looking odd triggered widespread allegations it was a deepfake. A week later the military attempted a coup.

In the lead-up to the 2020 US presidential election, deepfake parody TV commercials of Kim Jong-un and Vladimir Putin were commissioned by non-profit organisation RepresentUs to warn about the fragility of democracy. But RepresentUs says American TV networks refused to run the videos as they were too realistic.

For Ferraro, the nightmare scenario would be a deepfake video of US President Joe Biden declaring war on North Korea, potentially released at a time of heightened tensions on the Korean peninsula, with hackers posting the video on the White House Twitter account. Ferraro suggests such a video could create mass confusion in a matter of minutes, potentially triggering conflict and even casualties before it could be debunked.

If one person with a computer and an internet connection can make a convincing Tom Cruise, Ferraro fears the risks the technology could pose in the hands of well-resourced actors.

"Chris Ume said it took him a couple of months to train the algorithms and then he estimates about 24 hours to create a minute of video," he says. "But imagine if you're the intelligence services of China's People's Liberation Army, known as the 2PLA. They could put 10,000 man hours against the creation of a deepfake tomorrow."

Just as troubling are the implications of increasingly realistic deepfakes for the health of liberal democracy, and even our ability to know what is true.

Nina Schick, an author and governmental policy adviser who focuses on how digital technologies are reshaping geopolitics and society, says individuals' ability to create sophisticated deepfakes has "really shaken the political establishment".

The mere existence of convincing deepfakes is already "undermining trust in all authentic media", she says, and fuelling a conundrum called the "liar's dividend", where even real videos can be written off as deepfakes by those who deem them inconvenient.

"If everything can be faked, including video, then everything can be denied," says Schick.

A cat and mouse game

The potential for deepfake disruption is a problem the US government is taking seriously. 

Earlier this year, the FBI put private companies on notice to expect attacks from "synthetic media", meaning deepfakes, within the next 12 to 18 months, a warning Ferraro labelled "extraordinary". 

The US government, along with internet giants such as Facebook, are also investing in deepfake-detection systems. Almost every month now, a tech company claims to have made a breakthrough in beating deepfakes.

But many experts fear detection is a lost cause.

Facebook recently invited specialists to participate in a challenge to pit their detection systems against a series of fake videos. The winner from around 2,000 entrants was only able to detect a deepfake 65 per cent of the time, far short of the mark experts say is necessary.

"These deepfake detection tools, it's like a cat and mouse game," Ume says. "They can't follow. The deep Tom videos, they are pretty good. I'm not saying they're flawless, because I see a lot of mistakes myself. And they cannot detect it."

Without effective detection tools, some are now proposing to shift the focus to a system of digital "provenance", where legitimate videos are authenticated with a digital tamper-proof watermark guaranteeing their origin. 

Proponents of this system reason that if we can't prove what's fake, we should try to prove what's real by authenticating legitimate media as it's created.

Tech and media giants including Adobe, Microsoft, Intel, Twitter, the New York Times and the BBC are already collaborating on the world's first standards body on digital content provenance.

There are also moves in the US to regulate deepfakes with legislation, outlawing egregious uses of the technology in pornography and politics. 

Ume supports some form of regulation but remains a firm believer in the creative possibilities of deepfake technology.

"People are always scared of things they don't understand or they don't really know," he says. "I think it's a good thing I created these videos because now I'm raising awareness and they realise, 'This is real. It's coming.'"

Watch Foreign Correspondent's American Deepfake now on YouTube and iview.