Deepfakes are fake videos, aided by artificial intelligence (AI), that are so lifelike they become undoubtable. They are dangerous and, sometimes scary, but can the technology be used for good?
Elaine Burke, editor of Silicon Republic, joined Louise McSharry on RTÉ 2FM to explain what deepfakes are and how they're being used.
"The term itself is more used in the scary context of it," Elaine explains, "but I think a lot of us would actually be familiar with the style of technology because, like a lot of AI, it's actually already in our lives."
"There's a filter on TikTok that animates faces in photographs," she continues, "its a very similar concept in that you could kind of transplant someone else's expressions and facial movements onto someone else's look."
Check out this Bill Hader impression of Al Pacino where AI is used to create a deep fake video.
— 𝐌𝐨𝐡𝐚𝐦𝐞𝐝 𝐆𝐡𝐢𝐥𝐚𝐧 | محمـــد غيـــلان (@MohamedGhilan) August 9, 2019
If this is the technology available for the mainstream to make "fun" videos, can you imagine what governments have access to and how it can be used to manipulate populations? pic.twitter.com/jhVt4vzgL7
Much like how Bill Hader can school his features into that of Seth Rogan, Al Pacino or Jack Nicholson during his comedic impressions, deepfakes are a digital study of human expressions.
Elaine explains that the machine learning algorythm will look at numerous photos of a person, from different angles and lighting, so that it can map an image of their face onto another person with haunting accuracy.
"It does have that uncanny valley effect," says Elaine. A popular term among those who work in tech, uncanny valley references the breaking point at which people find technology unsettling.
"It will bottom out at a point where something becomes a little too looking like real but also not. It's when you see robotics that look too like people but you can tell that they're not people - that's really creepy for people. Deepfakes cause that as well when they're not to a very high standard."
When made to a high standard, however, deepfakes become unsettling for a whole other reason. They can become so real that it would be easy to believe that the person in the video above is in fact Tom Cruise rather than an impersonator.
"If someone can 'make' a very famous or noticeable face do what they want them to do, that can be used for political disinformation or misinformation," notes Elaine.
Celebrities are easy targets for deepfake material due because mass amounts of photos and videos are available at our fingertips.
Elaine says that making a deepfake of a member of the public would be more difficult, however, thanks to the popularity of social media, it certainly wouldn't be impossible.
"You can still make dodgy looking deepfakes with just a few photos and video footage of someone, and I think when it comes to pornographic material, people would be pretty upset even if it didn't look very convincing that something was real."
#Deepfake pornography is conventional pornography that has been doctored, often by #ArtificialIntelligence technology, so as to insert a celebrity appearing as (most commonly) the female actress starring in the scene, replacing the original porn performer.
— ConsentRevolution (@ConsentRev) August 7, 2021
Source: Vincent Barnett pic.twitter.com/TDqsVazed6
On a lighter note, this AI technology has been used for good in special effects on movies and TV shows.
In fact, a YouTuber who goes by the name Shamook was recently hired by Lucasfilm's Industrial Light and Magic following one of his deepfake videos going viral. The video saw him improve upon the VFX used in The Mandalorian where actor Mark Hamill has been made to look like his younger self playing Luke Skywalker.
"The people who are doing the best stuff actually work in visual effects in the media," she explains. "It takes a lot of work. The deepfake Tom Cruise TikToks are like 15 seconds long but they took weeks I'd say of work."
So, how can we spot a deepfake? Well, unfortunately, we might not be able to.
"By the time we have advice on how to detect them, the new software has figured out how to get around that detection," she says. "There are, similarly, well-equipped AIs that are made to detect deepfakes. Microsoft has built AI to detect them, I think Facebook has as well, but they're all struggling to keep up."
"The tells are often around the eyes," she adds. "Sometimes a deepfake doesn't blink quite just enough for a human being to blink. Sometimes the reflections in their eyes - their eyes just don't look real in a way, they do look kind of digitised."
"The hair often doesn't move quite right," she continues "You'll often see when impressionists do deepfakes that they'll use their own hair and just have the mapping on the face because - even animators will say this - hair is really difficult to get right as a fake."
To listen to the full interview listen back to Louise McSharry on RTÉ 2FM here.
