Using artificial intelligence, it is now possible to create fake photos and videos almost indistinguishable from real ones. These are called deepfakes.
To appreciate what’s involved with deepfakes, it’s useful to look at fake images historically. Consider the famous artist Michelangelo (1475-1564).
When he was young, Michelangelo created a sculpture in the style of ancient Romans, which his agent sold to Cardinal Raffaele Riario. When the cardinal found out it wasn’t authentic, he blamed the agent and gave Michelangelo an opportunity that helped launch his career. Centuries later, the actual sculpture was lost; this is one in the same style.
Forgeries have been a staple of the art world ever since. Talented painters can produce works in the style of past masters, and often only experts can tell the difference. For example, Charles Brooking (1723-1759) painted marine scenes. This sea battle, in the style of Brooking, was actually painted by Ken Perenyi, 232 years after Brooking died.
New technologies enable new forms of fakery. In the 1800s, photography became popular. In the mind of the public, photos were faithful representations of reality, and photographers took advantage of this assumption by manipulating images in various ways. This 1846 photo of four monks in Italy, the Capuchin Friars, was by Calvert Richard Jones.
Later, the negative was discovered. It shows a fifth figure (black in the negative) in the background, which has been removed.
Later still, it was discovered that the fifth figure was not another monk but a statue. Jones removed it to make the photo look less cluttered.
Consider this photo of General Ulysses Grant, US civil war hero and later president.
The photo is by Levin Corbin Handy, but it is not authentic. Produced in 1902, it was created from three 1864 photos. The head is from a photo of Grant standing.
The horse and torso are taken from a photo of a different general, Alexander McCook.
Finally, the background is from a photo of an internment camp.
Here’s a clue that the composite photo is not genuine: although Grant was a famous general, the soldiers in the background are paying no attention to him.
The Soviet dictator Josef Stalin tried to rewrite history. He employed a crew to retouch photos, often to remove individuals conflicting with Stalin’s preferred storyline of his regime. For example, consider this photo with Stalin in the middle. On his left, next to the canal, is Nikola Yezhov, who oversaw Stalin’s purges.
Later, at Stalin’s command, Yezhov himself was killed. He was also removed from photos.
For several of these examples, I have drawn from the fabulous book Faking It by Mia Fineman. It is filled with faked photos and explanations of their social context, and counters the common assumption that photos are a more realistic portrayal of reality than paintings.
Photos are routinely “retouched” to serve the desires of the photographer or the subject. With the arrival of the software program Photoshop, retouching became far easier. Retouching is regularly used to make people look younger and more attractive. Just putting “retouched photos” into a search engine gives numerous examples of faces before and after. Most of them show women, especially young women.
Some, though, are of men.
You might imagine that with the manipulation of images now being so easy, their credibility would be undermined, but a logical, rational assessment of a photo can be overwhelmed by an intuitive emotional response. When girls see pictures of models in magazines, billboards or social media, do they dismiss them saying “That isn’t real”? Or do they feel personally inadequate because they don’t measure up to a curated version of beauty?
A video is basically a sequence of photograph-like images, presented fast enough to give the impression of movement. It is technically possible to retouch a video by retouching each of the images, but that is labour-intensive. There are easier ways to fool people, especially people who are ready to be fooled.
One technique is to start with a real video and slow it down. The sound will also be slowed down, making it lower in pitch, but there is software to restore the original pitch. Consider a video of Nancy Pelosi, formerly speaker of the US House of Representatives.
An editor slowed down a video of Pelosi speaking, so she sounds as if she is drunk or disabled. This manipulated video was widely circulated, finding a ready uptake among those who disliked Pelosi’s politics.
Another technique is to splice together separate videos, giving the impression that they are part of the same event. Consider a video of Dwayne Johnson, known as the Rock, seeming to serenade Hillary Clinton.
If you look closely, you’ll notice that the segments showing Clinton are from a different event than the one where the Rock was singing. This type of video is called a shallowfake. It does not require advanced technology. Despite its obvious flaws, this particular shallowfake was widely circulated and, according to comments on YouTube, many viewers thought is was genuine. These viewers would have been confused when, during the 2016 US presidential election, the Rock came out in support of Clinton.
Artificial intelligence or AI is the ability of a machine to do things usually thought to require human intelligence. In recent years, AI capacities have greatly increased. AI can be installed in human-like robots, as in the case of Sophia, who became a citizen of Saudi Arabia in 2016.
One application of AI is chatbots, several of which have been released, most famously ChatGPT. Other applications include autonomous weapons, facial recognition, social media analysis — and deepfakes.
Imagine pitting an art expert against a forger. The art expert develops sophisticated techniques to distinguish a genuine original from a forgery, like the Ken Perenyi painting in the style of Charles Brooking.
But the forger learns about art-expert techniques and develops ways to produce more convincing fakes.
As this contest continues, both the art expert and the forger become more skilled, and the forgeries become ever better, detectable only by experts, and sometimes fooling them.
A deepfake is an AI-created image showing something that didn’t happen. It can be a photo or video. One way to create a deepfake is by using a “generative adversarial network” or GAN.
Two AI programs are pitted against each other, analogously to the art expert versus the forger. One AI program tries to produce a convincing fake and the other program tries to distinguish the fake from a genuine image. Unlike the art expert and forger, these AI programs can operate at great speeds, so the contest quickly results in a fake image that is hard to distinguish from a genuine article.
Consider this photo.
It was produced by a GAN, and is not an actual photo or even a modified photo. It was created, essentially, out of nothing. If you go to the website http://this-person-does-not-exist.com, you can see a GAN in action. When you refresh the page, within seconds the GAN produces a new face of someone who never existed. This is the digital equivalent of rapid-fire fake artistry in action, with realistic faces produced before your eyes.
This technology can also be used to create videos, for example this video of former US president Barack Obama seeming to say things he never actually said. Someone else speaks, and AI creates Obama’s face and voice saying those words, with appropriate accents and facial expressions.
Sadly, most deepfakes so far are porn. Using actual video of an actor speaking, AI is trained to be able to reproduce the actor’s face and position it on someone else’s body. One popular target is Scarlett Johansson.
Online there are hundreds of videos in which a porn star’s face is replaced by Johansson’s face, so it seems like Johansson is participating in the pornographic activities. The men who create these deepfake porn videos of course do not seek Johansson’s permission, nor the permission of the porn stars whose faces are replaced by Johansson’s.
Only a few years ago, considerable skill was needed to create convincing video deepfakes, but the technology has advanced so much that less know-how is required. This means that non-celebrities are more likely to be targeted, nearly all of whom are women. One woman subject to deepfake porn was Noelle Martin. She fought back, becoming a law reform activist.
Deepfakes can also be used for political purposes, such as a fake video of Zelensky calling on Ukrainian troops to surrender.
It didn’t have much of an impact; it lacked realism and credibility. But more sophisticated political deepfakes are possible.
Before raising the alarm about deepfakes, it’s important to recognise their possible benefits. For educational purposes, deepfakes can reproduce historical events, let’s say Hitler talking to his generals, providing greater understanding. Individuals who lose their voice, or who are seriously disfigured, might use deepfakes to communicate, with their actual words conveyed by a simulated version of themselves.
Another possibility is to enable a person to convey the same message in different languages, including ones they cannot speak. Manoj Tiwari, an Indian politician, authorised deepfakes of himself speaking in several different languages.
In this advertisement raising awareness about malaria, famous British footballer David Beckham seems to be speaking in nine different languages.
What distinguishes positive uses of deepfakes from damaging uses? A key consideration is consent. Manoj Tiwari and David Beckham agreed to the production of videos showing them speaking in languages they did not know, whereas Scarlett Johanssen and Noelle Martin did not consent to being in deepfake porn.
But what about those who are dead? We may not care about obtaining permission from Hitler and his generals, but would you be pleased if, after you died, a deepfake of you was used in a film? It might depend on the film or the role. Maybe it would be okay if your face replaced that of Julie Andrews in Mary Poppins.
On the other hand, you might not be pleased for your face to replace that of actor Anthony Perkins in Psycho.
With the advent of technology that makes it easy to make deepfakes, many people are potentially at risk. Is there a video online of you speaking for a minute or more? If so, this video could be used to make a deepfake video. Even a single photo, in someone else’s hands, can be used to create a deepfake photo. If the deepfake shows you in a compromising situation, it could be used for blackmail or humiliation. If you’re not a celebrity or don’t play a prominent role in some enterprise, your risk of being targeted is probably low. But Noelle Martin’s experience shows anyone can be attacked. So maybe you should prepare, just in case.
There’s another implication. As more people become aware of deepfakes, genuine photos and videos may be questioned, claimed to be deepfakes. This is especially important for groups trying to expose human rights violations, using images to raise awareness and stop harmful practices, like New Tactics in Human Rights and Witness. If pictures of atrocities can be dismissed as fakes, the efforts of human rights campaigners will be more difficult. We need to be prepared for loss of trust.
There is a lot of writing about deepfakes. Much of it is highly technical. Even social analyses are mostly found in academic journals. The most entertaining treatment I’ve found is the book Trust No One by journalist Michael Grothaus, who was able to contact some of the producers of deepfake porn, hearing their side of the story.
For a scholarly yet accessible treatment, see Graham Meikle’s book Deepfakes. Meikle’s discussion led me to Mia Fineman’s book Faking It, about fake photography before Photoshop.
If you look online, there are ever more clever videos showing the power of deepfake technology, some of them providing warnings about implications. Here are a few that I found revealing and entertaining.
Former US president Richard Nixon reads a script prepared for him to use if the first men on the moon had not come back. Fortunately, they returned safely, but with this deepfake we can see what the world would have seen if history had been different.
Deepfake of Kim Joo-Ha, news anchor for Korean TV channel MBN.
I’m especially interested in what peace and human rights activists should be doing to prepare for a future in which deepfakes are more common, and potentially used against them. Let me know if you have suggestions.
Brian Martin, firstname.lastname@example.org