Is called Hany Farid and is considered one of the world’s leading experts on deepfake videos. This digital forensics expert was capable of uncovering videos manipulated by governments, for example, but now he has decided to leave Silicon Valley for one simple reason: It is no longer able to differentiate the real ones from those that are being generated with AI tools. And we are not surprised.
Deepfakes indistinguishable from reality. In the last two decades Farid, 60, has specialized in identifying fake videos. This professor at the University of California at Berkeley has confessed that advances in generative AI have made traditional detection methods are no longer of any use. Their conclusions confirm the feeling that we have had for a long time with this type of content: AI has advanced so much that the problem is no longer just deepfakes: it is that we distrust even real photos.
Farid’s reputation precedes him. His father worked for 50 years as a chemist at Eastman Kodak, and Farid grew up visiting the dark room often, watching photos become photos as they passed through the different liquids. He ended up designing a “digital fingerprint” system that made it possible to detect cases of child pornography hidden on the Internet. In fact, its technology has led to 30 million cases of potential abuse being reported each year, as well as leading to hundreds of arrests and several rescues.
I surrender. Faced with the avalanche of perfect deepfakes generated with AI, Farid has decided to leave his job to take refuge on a farm in Vermont. His surrender is the latest demonstration of a harsh reality: We can no longer trust what we see on networks. Now he is dedicated to working with wood, and has distanced himself from networks and technology.
The missile that changed everything. The turning point that demonstrates this crisis of this digital forensic task occurred after the viral spread of a video showing the alleged impact of a US missile on a school in Iran. Farid spent an entire day breaking down the sequence frame by frame: analyzing the geometry of the shadows, the sound delay rate of the explosion according to the laws of physics, or the pixel length of the projectile.
Impossible to decide if it is false or not. He found nothing that could prove that the video was fake, and the same thing happened to other specialists. None could issue a clear verdict of authenticity, and that made it clear that AI video generation is currently so advanced that real content is indistinguishable from a deepfake generated with these latest generation models.
Verifying is too complicated. There is another problem here: generating a fake video, whether toxic or not, with cloned voices that are perfectly synchronized with the interlocutor is easy, fast and cheap. Carrying out a forensic investigation to try to detect whether the video is real or not takes hours of computational and direct analysis by specialists. Given that deepfakes manage to go viral in just 20 minutes if they are successful, the methods to contain this spread are useless for a simple reason: they arrive late.
The biter bit. The researcher himself was a victim of this reality: cybercriminals cloned his phone number and used AI to generate his voice and thus impersonate his identity. With that clone, they called a close contact who was involved in a court case and managed to extract confidential information. Farid and his wife, a vision researcher at Berkeley, they had to create a secret safe word at the beginning of each family call to certify that each interlocutor was who they said they were. The situation generates a disturbing paranoia and mistrust.
“I’m going blind”. In the report of The New York TimesFarid explained that his studies show that most people can no longer differentiate a real photo from a digitally created one. “I feel like I’m going blind,” he indicated, showing his concern about an AI that is managing to obscure the truth and distort reality.
Watermarks as a solution. Faced with this avalanche of images and videos generated by AI that are indistinguishable from reality, one of the potential ways to mitigate the problem continues to gain strength. It is, of course, the watermarkstotally invisible and which are part of the metadata of those files.
Two promising initiatives. There are several initiatives in this regard, although the most notable It is that of the C2PA coalition which includes, for example, Google and OpenAI. AI tools should add those watermarks identifying those contents (“This video has been generated with this AI application, this image has been generated or edited with this other one”), but at the moment that type of option is not applied by default. Another important project in this sense is SynthIDGoogle’s technology to “mark” these contents as created with AI.
GIPHY App Key not set. Please check settings