Trump and Biden debate - or is it?

Deep diving into deepfakes

Image credit: Dreamstime

In ‘Trust No One’, Michael Grothaus argues that AI-enabled deepfakes have moved on from amusing digital manipulation to something more sinister.

“The weaponisation of deepfakes against politicians or nation states has become something we’re simply going to have to live with,” says Michael Grothaus. Author of ‘Trust No One’, an investigation into the nature, origins and future of the deepfake video. Grothaus thinks that the days of innocent ‘face-swapping’ for the amusement of the YouTube audience, or even the merging of celebrities into pornographic videos, have dramatically transformed into “a real threat.”

When the public can routinely view plausible videos of events that haven’t taken place, globally distributed across social media channels, he says, “you start to see the erosion of trust in society. People will become more cynical and will start to think that everything they see is fake. So it is more essential than ever to understand the dark origins of deepfakes. The more deepfakes spread, the less we will be able to tell if the videos we are watching are authentic.”

There was a time when fake news was a non-factual item of written journalism, says Grothaus, who is also the author of the novel ‘Epiphany Jones’, which examines alienating aspects of the internet in the Hollywood sex-traffic trade. That was when the impact could be controlled by news editors and damage-limiting press conferences. Fake video is another story: “Unlike these shallowfakes, where the story is manipulated without sophisticated machine learning, the deepfake harnesses the power of artificial intelligence.”

Grothaus is keen to stress that the word deepfake has a specific meaning related to video. It doesn’t apply to still photographs, which have been subject to manipulation since the invention of the camera. A portmanteau word, the ‘fake’ component, says Grothaus, “is self-explanatory – its output is fabricated – while the ‘deep’ descriptor may not seem so obvious. It refers to deep learning, a type of machine learning which is, of course, a type of artificial intelligence. The deepfake is a video that has been manipulated by AI, and the way it works is great, but it’s also terrifying in that it shows you how powerful AI is.”

We read it for you

‘Trust No One’

AI has enabled the evolution of the counterfactual video – the deepfake – that allows us to see the world not as it is, but how the videographer wants us to. While once the idea of manipulating photos was a largely innocent pastime done for amusement, creating realistic fake videos, says Michael Grothaus in ‘Trust No One’, has far wider-reaching social and political implications. It might seem harmless to watch the United States football team raising the World Cup, but there are few barriers to deepfakers producing clips showing the President of the United States declaring nuclear war on China. The deepfake’s biggest ally in spreading misinformation of this type is the amplification effect of social media. The only limit on deepfakes in a post-truth world is the imagination, which is why we need to be able to identify them, says Grothaus. Chilling stuff.

At this point Grothaus explains how the deepfake pits two types of AI against each other simultaneously while viewing the same set of media. On the one hand you have the ‘forger’ and on the other you have the ‘inspector’. The forger AI reviews authentic photos of the face that is being deepfaked, copies them and presents them to the inspector, “that will then evaluate whether the image is real or obviously fake. The first few thousand attempts by the forger to create a deepfake will result in something that looks pretty bad: eyes in wrong place, nose crooked, pixilation. But every time the inspector rejects an image, the forger learns and tries to forge a better version. That’s the magic of AI here. The forger is training itself to get better at making images.” After these iterations, the forger is then able to create pictures that are indistinguishable from real imagery.

This is where the deepfake enters murky ethical territory. While there is endless scope for harmless deception – maybe you’d like to see yourself on stage with the Rolling Stones – there is also huge potential for large-scale societal deception and damage. At the start of ‘Trust No One’, Grothaus paints a counterfactual of the run-up to the 2020 US presidential election in which a video gets posted on Facebook depicting Joe Biden stating that when elected, his first action will be to sign an executive order confiscating all firearms. The video is what the cyber experts had feared. But there’s no need to worry because it is an amateurish fake and the discerning electorate will realise that.

The problem is that the video has by now been disseminated over social media and, convinced that the deepfake is real, thousands of ‘undecideds’ among the pro-gun lobby turn out in force to vote for Donald Trump. Elsewhere in the book we have a scenario in which Trump is seen on video declaring nuclear war. While neither of these events happened, “the problem the deepfake video brings is its amplification on social media. Lies can travel halfway around the world before the truth has even got its boots on,” says Grothaus.

‘What we need is more education. We need to make viewers aware that not every video they are watching is real.’

Michael Grothaus

One of the key effects of deepfakes in society is that it skews what is believable: viewers will lose the ability to differentiate between authentic and fabricated news. But there is the further effect of providing criminals and celebrities with undue deniability in the legal process. Standard Sunday tabloid fare, such as a footballer getting caught on video with a prostitute, now comes literally with a get-out-of-jail card: “The guy just says it wasn’t him. The video was a deepfake.” This type of deniability is termed the ‘liar’s dividend’: miscreants can simply denounce authentic evidence as fake. The societal problem is that the dividend (or benefit of the doubt in favour of the criminal) grows in direct proportion to the public’s awareness and suspicion of deepfakes.

Back in the day, says Grothaus, “if I recorded you doing something horrible – such as setting fire to a cat – on video, that would be enough for me to turn you in for animal cruelty. But we’ll soon be at the point where you can defend yourself by saying: ‘Hey, that’s not me. I know it looks like me, but it’s a deepfake’. This is the downside the technology brings.” It’s an issue we have to take seriously, says Grothaus, because “people have the right to defend themselves, especially when they feel they are being framed by the clever use of technology. It’s another dimension of how deepfakes will affect us all.”

Which leads to the issue of what can be done to establish authenticity in a post-truth environment. “That really is the question,” says Grothaus. “I’m a great believer in technology as a force for good. I also think that deepfake AI technology is neither good nor bad. Creating legislation to outlaw the technology itself is the wrong way to go.” Maybe you could legislate about the way it is applied, he suggests. But the problem is that the sort of criminals who generate celebrity deepfake porn are hardly likely to be deterred by being told not to do it. “What we need is more education. We need to make viewers aware that not every video they are watching is real.”

‘Trust No One’ by Michael Grothaus is from Hodder & Stoughton, £18.99


Future of film?

Given all the opportunities deepfake technology opens for Hollywood, it’s hard not to get excited about how the tech could usher in the next era of filmmaking. Deepfake technology could be as transformative to the film industry in the 2020s as sound was to the 1920s, colour cinematography was in the 1950s and computer graphics were in the 1990s.
Right now, most deepfakes have one major technical limitation: the deepfaked faces can’t be any larger than 256 pixels high by 256 pixels wide. This resolution is adequate for the deepfaked faces appearing in the fancasts posted to YouTube and in fake celebrity porn because those videos are typically viewed on small screens. But [at the] cinema, the deepfaked faces would look blurry. They’d be too low-res compared with the sharpness of the rest of the massive picture.

So, what’s Hollywood to do if they want eventually to be able to make billion-dollar blockbusters with the world’s hottest actors for the rest of time? Well, they need to work on improving the quality and size of deepfakes themselves. And that’s just what computer scientists from Disney Research Studios have been working on. In 2020, they announced that by using new neural algorithms and other advanced machine-learning techniques they were able to increase a deepfaked face’s resolution sixteen-fold. As the company’s researchers showed, Disney’s deepfakes look sharper, more detailed and have no discernible seams or artefacts. Of course, Disney’s researchers aren’t developing their deepfake tech to help seedy deepfakers create higher-quality fake celebrity porn. The endgame is to make better films with more realistic synthetic imagery in much less time than traditional, costly, CGI techniques take.

Edited extract from ‘Trust No One’ by Michael Grothaus, reproduced with permission.

Sign up to the E&T News e-mail to get great stories like this delivered to your inbox every day.

Recent articles