What Are Deepfakes?

Today, you can put your favorite celebrity in a fake situation or put any face of your choosing on top of any body. You’ve probably seen the phenomenon all over social media, with opposing presidential candidates kissing or celebs fighting, or worse.

Deepfake technology poses a threat to news media the likes of which humanity has ever seen. If the spread of misinformation was potentially destructive, one can see how deepfake videos can cause catastrophe. So, if you thought fake news was bad, deepfakes just made everything much worse, so lets learn all about how bad things can get.

Table of Contents

• What Are Deepfakes?

• The Problem with Deepfakes

• How to Spot Deepfakes

What Are Deepfakes?

Deepfakes utilize artificial intelligence with deep learning to generate images or audio of fake events. Deepfakes are used for memes, propaganda, and deepfake pornography, with over 96% of deepfakes being pornographic in nature.

Using machine learning from datasets of people’s faces, bodies, and more, these AI systems can make face-swaps by having an algorithm run thousands of images through an encoder to find the similarities between the facial features and facial expressions of two people, then compresses the image to only each face's shared similarities. The algorithm then runs the compressed image through a decoder to reconstruct the desired face and swap out the other one.

Audio recordings are also the subject of audio deepfakes where AI technology takes a familiar voice and features it saying any script a user inputs.

The Problem With Deepfakes

AI-generated deepfake content opens up the potential for fraudsters and cyber criminals to engage in defamation of any public figure, scams, phishing, impersonation, and identity theft of real people.

One notable case involved a finance worker from a multinational firm that was tricked into paying out $25 million dollars to a fraudster using deepfake technology to pose as the company’s CFO in a video conference call.

Deepfakes pose a deeper problem to geopolitics than simply a threat to cybersecurity vulnerabilities. Being able to generate videos of world leaders saying anything you want may be easy to identify today, but someday undetectable AI versions of deepfakes may trick the world into thinking all sorts of false information or disinformation.

For example, London mayor, Sadiq Khan, was the subject of a deepfake audio recording of him making inflammatory remarks before Armistice Day. The audio, which featured calls for protest on a national remembrance day, caught national attention and nearly led to major disorder.

Sexualization of public figures is also a major issue, as thousands of these images are being distributed across social media platforms like Facebook and Instagram. Meta's Bullying and Harassment policy prohibits "derogatory sexualized photoshop or drawings" on its platforms, but still, Mark Zuckerberg has had difficulty curbing deepfakes.

Elon Musk has seen his own share of deepfake problems, although pornography isn’t prohibited by X’s content policy, his platform had to temporarily restrict searches for Taylor Swift to stop the spread of deepfakes.

The impact of Deepfake pornography hits the victims hardest, as images that breach their privacy proliferate without their consent, they can feel isolated, depressed, and unsafe online.

How to Spot Deepfakes

Many AI generated photos, especially in the early days of the technology, had many obvious giveaways. There’s a certain electronic sheen to AI generated photos.

Like text generated by AI, AI detectors exist to scan the content for the watermarks of AI. Deepfake detection, like AI plagiarism detectors, can easily be bypassed the same way undetectable AI services are programmed to do. Deepfake detection checks for the watermarks of generative AI, so making undetectable deepfakes would simply replace those watermarks with the watermarks of humanness.

Still, maintaining a critical eye when viewing an image is key to detecting deepfakes. Inspecting anatomical distortions, differences between skin tone of face and body, inconsistency between light and shadows, sharpness of the edges of a face, and if it's AI video speaking that you're inspecting if the AI speaking's teeth are sharp or blurry.