• The new frontier of hoaxes and fake news is deepfakes
• Artificial intelligence risks giving rise to urban legends that are hard to dispel
• Some solutions for protecting yourself against this drift in information

"What is the perception of reality? Is it the ability to capture, process and make sense of what our senses receive? If you can hear, see, touch or smell something, does that make it real?" These were the words spoken in a video posted on YouTube last year by Morgan Freeman, one of Hollywood's most enduring and famous stars. Except that, at the end of the clip, Morgan Freeman himself declared that he was not Morgan Freeman: what our senses as viewers had just received was a deepfake — that is, a new application of artificial intelligence and machine learning capable of blurring the boundaries between truth and falsehood.

What is meant by deepfake?

The term has rapidly colonised the media and much of the new collective imagination, but what are deepfakes really? To begin with, it should be noted that, despite the world only starting to talk about them in the years surrounding the Covid-19 pandemic, deepfakes existed — and were advancing — for nearly a decade before that: in 2014, Ian Goodfellow, formerly head of Apple's artificial intelligence division, published a scientific paper introducing the concept of a "generative adversarial network" (or GAN), a particular application of machine learning in which two neural networks "compete" in a zero-sum game that enables them to learn to generate new images from an existing dataset. And in 2017, researchers at graphics card giant Nvidia enhanced and made the results of GAN networks more convincing.

This was the birth of the deepfake: websites such as thispersondoesnotexist.com made low-cost falsification tools available to the general public on a scale unprecedented in human history, and for the first time the world realised that our relationship with reality was hanging by a thread. "Deep" because the neural networks behind these generation tools are made up of multiple layers: this is deep learning, that form of artificial intelligence that makes it possible to replicate the appearance and voice of real people; "fake" because, well, that needs no explanation.

In November 2017, an anonymous user posted on Reddit an algorithm capable of producing fake but realistic videos: the code ended up on the popular code-sharing site GitHub, where it was made freely available to anyone. And from that point on, deepfakes found their way into the first major sector in which they would come to prominence: pornography. Applications such as FakeApp allowed anyone to insert any person's face into compromising videos that at first glance appeared realistic, generating an endless number of risks and a great deal of misogyny: Scarlett Johansson, who unwillingly became one of the most frequent subjects in fake pornographic videos circulated across the internet, said that fighting them was "a lost cause," since "the internet is a dark black hole that feeds on itself."

What are the most famous deepfakes

Since its early days, the "deep fake" has come a long way, and some of its manifestations have led public opinion to question the merits and risks that this new technology brings with it: in addition to the already-mentioned Morgan Freeman deepfake video, the work of Dutch creators Bob de Jong and Boet Schouwink, mention must at least be made of the fake surrender of Ukrainian president Volodymyr Zelensky, circulated by Russia last year, and the deepfake of actor Tom Cruise that went viral on TikTok, created by special effects expert Chris Ume and Miles Fisher from hundreds of thousands of photos and videos of the Hollywood star.

Deepfakes also made their appearance on television, further expanding the audience exposed to their illusions: across the Atlantic, the satirical programme Saturday Night Live broadcast in prime time a deepfake of Hillary Clinton in tears that was entirely indistinguishable from a genuine video, while in Italy Striscia la Notizia did the same with an algorithmically generated video of former prime minister Matteo Renzi. In 2018, the online outlet Buzzfeed had circulated a fake-but-convincing clip of Barack Obama, which only revealed in the final moments that it was performed by actor Jordan Peele.

The risks of deepfakes

With great power comes great responsibility, as a well-known comic book superhero once said, and the same applies to the unknowns looming on the horizon of deepfakes. What happens when thousands — and potentially even millions — of people believe a fake video of a political leader announcing a nuclear war? And before that: what will become of our relationship with reality, now that what we see and hear is no longer "true" by definition?

Even closer to home, at the time of the Matteo Renzi deepfake on Striscia la Notizia, many observers pointed out that the somewhat ambiguous presentation of the content in prime time could give rise to misunderstandings with serious consequences: Wired had described it as "a further challenge for journalism and for public opinion as a whole."

Julia Bayer @bayer_julia

𝕏

How to spot a deepfake? Let's say: spot the (in)visible 👁️🔍
Together with @RubyBouw we compiled hands-on tips: dw.com/en/fact-check-… Thread

♥ 102💬 4 21 gen 2022 · Apri su X →

But abroad, as is often the case, the debate on deepfakes is far ahead of what makes the front pages in Italy: last year, for the first time, a large-scale campaign of videos showing fake television news anchors with an unusually pro-Chinese slant flooded English-language Facebook and Twitter, opening a new and worrying front in geopolitical information warfare.

Who will save us from the invasion of ultra-fakes?

Certainly not legislation, at least for now: the only state to have passed urgent measures in an attempt to curb the phenomenon is China, which at the start of this year made it mandatory to obtain the consent of the person being "faked" and to display distinctive markers — such as watermarks — attesting to the fictitious nature of the work. But even the dragon faces the same problems as the rest of the world: how do you rein in a phenomenon whose most rampant abusers often operate anonymously, on fluid platforms where content multiplies so rapidly that they can act almost undisturbed?

Something is at least beginning to stir in old Europe too: the Artificial Intelligence Act, just passed by the EU Parliament, aims to impose a new international standard for the regulation and transparency of artificial intelligence applications. But many parts of its text are the subject of debate and legal uncertainty: it is difficult to define the scope of what can be labelled "artificial intelligence," let alone how to regulate it. For now, then, those photos and videos of people who do not exist are set to do as they please — and we, as in that great Sanremo classic, will no longer be able to see reality.

How to recognise a deepfake

In the face of all these risks, all that remains is to understand how to protect yourself from deepfakes — which is no simple matter, and will become increasingly less so as these technologies advance exponentially (in the latest version of Photoshop, Adobe has introduced a generative function capable of creating photomontages that are virtually indistinguishable from reality).

At the time of writing, the leading artificial intelligences still have some trouble with rendering human hands: if you see an upper limb termination that looks slightly off, do not rule out the possibility that it may be a fake.

MIT recommends, among other things, paying close attention to the facial skin of the subjects in suspected deepfakes: if it appears too tight, or conversely too wrinkled, or shows different degrees of ageing in different parts of the face, then there is a good chance that the image or video in question is not real. The blinking of eyelids and the movement of lips can also offer useful clues in this regard.