The World of AI Deepfakes is Coming

How can we survive if we don't know what's real?

I’ve had floods – biblical levels – of emails asking me to talk about AI and deep fakery, so here it is. If you have any specific issues you’d like me to attempt to think about (no actual thinking promised): nick@podotpods.com

I have been feeling very nervous about artificial intelligence. This is not because it is currently in anything like a state where we should be anticipating radical changes to the way we work or live. At present it is gimmicky and only semi-cooperative. But it should be offering a glimpse of where technology is heading. It is, I imagine, what it felt like when the World Wide Web was being mooted in the 1990s – even though, at the time, it was just a few academic super computers linked to a local network, there was a glimpse there of what would become the internet. And that glimpse of the internet was also a glimpse of our collective future, as a species.

AI doesn’t scare me. I don’t fear the robot uprising. I think it’s pretty much an iron rule of technology that you should minus 50% from your worst expectations for the future. The idea that we were going to slip, Tron like, into a digital space, has been, largely (I’m looking at you, Zuckerberg), debunked. The internet, now in its mature phase, is still to a large extent just cat videos and pornography and reviews of dry cleaners.

The reality is that the use cases for the sort of incipient AI capabilities we’ve seen in recent months are fairly slim. There, but slim. And technology is like panning for gold in reverse: only where there are real gaps will the stuff stick. The internet solved a lot of problems. Problems in communications, problems in knowledge distribution and acquisition, problems in creativity, and, yes, problems in the worlds of cat videos, pornography and dry cleaner reviews. AI will doubtless solve a lot of problems – we have already seen the way that ChatGPT, in a few short weeks, has finessed the search engine model that had been predominant for decades – but it will also run into user fatigue. The average person is not going to spend much of their life generating complex image or video models of, say, Boris Johnson unfeeling his skin to reveal a smaller, Matryoshka doll, Boris Johnson within.

There will, however, be plenty of sensible use cases. Triaging patients in understaffed hospitals, for example. Language tuition for those without access to group conversations. Optimising the very rudimentary world of virtual assistants, like Siri and Alexa. But I don’t think it will gazump the work of, to pick a totally random example, journalists. Because AI can undoubtedly replace the most disposable of journalism, but this is a bit like worrying about the impact that covid-19 might’ve had on other vicious respiratory illnesses. If your job, as a “journalist”, was rehashing press releases and wires from the Associated Press then, yes, your job might be under threat. Otherwise, I think you’ll survive. (And let’s not forget that for the past several years, many outlets internationally have already been utilising AI capabilities in their publishing…).

Because, ultimately, it’s not in humanity’s interests to destroy the world. There’s this nebulous political idea known as Fully Automated Luxury Communism; a world in which we will all live like kings while the robots and computers do our bidding. No work, all play. Except the horrific issues with supply chains and resource scarcity has rather poured cold water on this ambition. Can we mine enough rare earth minerals to forge the chips that will power our robot butlers? How big will the robots have to be in order to build (and maintain) the nuclear power plants necessary to warm our homes in a post-fossil fuels world? Why are we even bothering to humour this pretence?

It is not in humanity’s interest to create technology that destroys livelihoods but cannot transmute the fundamental issues of finite resources. And another reason to be optimistic about the AI non-takeover: one of the jobs most patently impacted by the rise of AI is the software developer. Already, ChatGPT can write code (albeit plagiaristic code) that would shame most Chinese high school students. I think the software development community can be relied upon to not do themselves totally out of a job.

And so I suspect that just like the reverse panners in the reverse California gold rush, it’ll only be the gaps through which the mud, sand and silt floods. Gaps in healthcare, social care and other welfare opportunities; gaps in education; gaps in logistics.

Then, separately, there’s the question of deepfakes. Deepfakes are not new, they go back to the Cold War and earlier. I’m sure that historians could source examples from Ancient Greece of fakery used for disruptive political purposes (“Yes Brutus, that’s definitely Caesar’s handwriting calling you flagitium hominis on the wall of the Agora…”). But AI has undoubtedly empowered the possibilities here.

Last week, an image emerged of Pope Francis – an elderly man currently convalescing with bronchitis – wearing a Balenciaga puffer. The internet went wild for the old man in the milk white floor-length Michelin man jacket. The Pope, they all said, has got quite a sick style. And it made a sort-of visual sense, because the Pope does dress in quite a blingy manner (most Roman Catholic clergy do – just look at the Cardinals’ natty hats). And the Popemobile, a white converted truck used to carry the Pope through crowds on world tours since the 1970s, has become a symbol of the Holy See’s descent into modernity.

The fact is, the Balenciaga Pope picture was a fake. A convincing fake, apparently cooked up by AI. It ticked all the key boxes for effective misinformation: i) it was good quality, ii) it was plausible, and iii) there was no obvious reason for it to be faked. Because of this combination of factors, people bought it.

This is, however, a pretty rare triumvirate. On the first point, AI is undoubtedly going to improve the quality of deepfakes. I could, with the free resources available on the internet right now, cook up a fairly convincing Donald Trump mugshot, within a few minutes. A couple of years ago, that would’ve required skills with Photoshop that I don’t currently have. The second point – the plausibility – is also key. If the Pope had been shown, I don’t know, gaffer taping the mouth of a whistleblower, it might’ve made a more powerful political point, but I would’ve immediately dismissed it as fake. So the search for plausibility, a vital part of misinformation, is quite a limiting one.

And then there’s the third point, the sort-of apolitical pointlessness of it. The Balenciaga Pope was an example of what would be called “shitposting” on Twitter. Anarchic, memetic, victimless.

An AI faked Donald Trump mugshot would be an interesting test of these principles. Certainly it could be good quality. A mugshot of Trump exists, somewhere in the New York criminal justice system, and therefore its publication is plausible. But it falls, somewhat, at the third hurdle. Both the pro and anti Tump sides have their reasons for wanting to publicise a Trump mugshot, and therefore its sudden appearance is likely to provoke scepticism and sourcing. And so it might fool some people – maybe even a lot of people – for a short while, but quickly that deception is going to be unravelled. The pressure on the question of reality is too great.

This is the key question in the world of AI deepfakes: can they be disproved?

If a Donald Trump mugshot appears and the NYPD release a statement saying it’s fake, I think that a degree of consensus would emerge fairly rapidly. After all, no particular interests are served by its existence or non-existence. If, however, a photograph of Donald Trump – I don’t know – being sprinkled with apple juice in a Moscow hotel room emerged, and the Trump camp called it a deepfake, would that satisfy all of the liberal media? Would the conservative media believe that a video of Joe Biden French kissing a… puppy was a deepfake, just because Joe said it was?

It is clear that AI is going to be used to consolidate bad opinions in the coming years. It will entrench political division because the average internet consumer is not very discerning. Let’s remember that humans – apocryphally, perhaps – screamed with fear at the sight of the first filmed train coming towards them in a Parisian cinema. And I don’t doubt that these same humans were baffled in the early days of the Adobe Creative Suite, when Photoshop suddenly made it easy to vanish (or implant) people and objects from places they’d never been.

And we do live in a world of pronounced, if not deep, fakery. Airbrushing is totally ubiquitous: from spots to waistlines. We live in a world of CGI, where TV shows and films aren’t set against elaborately painted backgrounds but against easily rotoscoped out block colours. Everything on TV, on Instagram, on TikTok, on cinema screens and Netflix and YouTube and Steam – they all involve a degree of fakery. The permeation of this into the public and political discourse was inevitable. Ultimately, our simian brains are still screaming at the approaching train.

But there is going to be a deluge of stuff that does satisfy my three criteria, and it’s going to be very hard to filter those. And I am not just concerned about the proactive damage of deepfakes, but the disruptive possibilities that the destabilisation of trust in hard evidence causes. I asked a lawyer friend the other day if there had been any discussion at his firm of what would be said if someone disputed materials presented in evidence, claiming they were deepfakes. He said that this was not a discussion that had ever happened.

So we may find ourselves sleepwalking into a situation where disreputable evidence is being introduced to the legal system, but, equally, the claim that evidence has been faked is very hard to disprove. This isn’t the Balenciaga Pope; this is an existential threat to an evidence-based rule of law.

And part of the issue is that the responsibility for creating detection methods for deep fakery is going to fall to the same men and women who cooked up the offending technologies. This isn’t just marking your own homework, but acting as foreman (and the other 11 jurors) at your own trial. And while I trust that software developers will be cautious about mainstreaming software that puts them out of work, I don’t trust them to act for the benign good of the world. I don’t trust lawyers, I don’t trust the police, and I definitely don’t trust technologists.

And so when I see things like the open letter, signed by technology luminaries like Elon Musk, Apple co-founder Steve Wozniak, and cognitive scientist Gary Marcus, calling for a pause on AI research, I feel a little bit more optimistic. Not because I expect any material impacts as a result of the letter (let’s not forget that Musk previously co-signed an open letter, with none other than Stephen Hawking, in 2015, asking for more research into the impacts of AI, which seems to have been roundly dismissed) but because I think these conversations have to happen within the technology community. They created the problem, they’re going to be asked to solve it; they need to be reckoning with it.

Because AI has been an innovation with, to date, too little reckoning. Passively accepting the inevitability that the worst habits of digital civilisation are about to be hard coded into the DNA of our collective future feels like madness. Deepfakes are not new – Stalin, after all, used to airbrush dissidents out of photos as easily as they were airbrushed out of, er, life – but we are about to unlock some revolutionary tools. In the decades since the invention of the camera, of sound recording, of video, those forms have become fundamental, not only to how we explore the internet, but to our modern empiricism. And if you shake that – destabilise it, to use the parlance of our times – there will be consequences. Not everything is as easily proved, or disproved, as the Pope wearing a Balenciaga puffer jacket.

And if there’s one thing I believe, at my core, it’s that we can’t leave it to technology to solve the problems created by technology.