What is a deepfake?

Deepfakes are AI-generated media that appears to be of real people saying or doing things they never said or did. They're almost always created to deceive viewers.

Back in the early 2000s, a software company called JibJab created interactive ecards where you could put anyone’s face in an animated video. It used a rudimentary AI face-swapping algorithm to show elves singing Christmas carols. They were entertaining, and cartoonish enough to avoid fooling anybody. While the idea is the same, modern deepfakes are much more sophisticated (and possibly harmful) than these JibJab ecards. 

Some well-known examples of deepfake videos are Hollywood crossovers—like Jim Carrey’s face replacing Jack Nicholson in The Shining. There are also more serious ones, like Jordan Peele’s public service announcement to warn the public about deepfakes and fake news using President Barack Obama’s likeness. 

How are deepfakes created?

Unlike Photoshop, deepfakes require machine learning and datasets to achieve a believable result. While that might sound complicated, creating a deepfake is actually easier than a lot of post production video editing techniques.

Open-source apps can make synthetic media for you with just a few clicks of the mouse. FaceSwap, DeepFaceLive, and DeepFaceLab have learning tools built in. You can even train your own face models to make deepfake content more accurate.

How do deepfakes work? Deepfake technology 101

There are a number of ways that deepfake technology operates. But essentially, an artificial intelligence tool called a neural network is trained to recognize the facial movements of a person when certain sounds are created. 

Using data from the neural network, a deepfake program creates a 3D map so it can match skin tone, facial expressions, and lip movements. It uses generative AI to make videos that appear like the face it has studied, and say whatever you want it to say. 

To make the deepfakes even more realistic, some use an alternative machine learning technique known as a generative adversarial network (GAN) to detect flaws and improve the quality of the deepfakes. 

With a dual-network architecture, generative adversarial networks (GANs) train two models at once: the generator and discriminator.

• The generator creates synthetic images or videos, trying to make them as realistic as possible.
• The discriminator evaluates these outputs, identifying differences between the generated content and real images or videos. 

This competition drives the generator to continuously refine its outputs, reducing the imperfections that make deepfakes recognizable.

🎙PRO TIP: If you’re going to create AI-generated audio (such as a deepfake of your own voice for your podcast), use a program that protects the rights of your voice and likeness. Programs like Descript—specifically its AI Voices feature using Lyrebird—require you to verify your voice before you can clone it. 

Are deepfakes dangerous? 4 risks you should know

Deepfakes don’t have to be dangerous, but so far, most deepfakes on the internet aren't consensual and are being used illegally. Here are four common ways deepfakes can be harmful:

Misinformation and disinformation

Misinformation refers to the false or inaccurate information that is spread around the internet. Disinformation is deliberately misleading or biased information, manipulated facts, or propaganda used to perceive people. Many studies have documented the rising tide of misinformation, and most conclude they’re a severe threat to public interests. 

One example: During Donald Trump's arraignment in April of 2023, deepfake photos circulated showing him being violently assaulted by police. This inspired sympathy and outrage from his supporters. But it wasn’t real. 

Whether deepfakes are political or not, the misinformation they spread can be dangerous and cause confusion about important issues. It can also delay reactions during time-sensitive events.

Legal consequences

Laws being created around fake content deal with privacy, copyright, libel or slander, extortion, and identity theft. 

As of now, most deepfake legal cases are civil, and the perpetrators won't go to jail. But more and more states are beginning to pass legislation on deepfakes, including California, Texas, and Virginia.

Reputation damage

Deepfaking hasn’t been used for good. Its use is mostly nonconsensual, like deepfake pornography with celebrities' faces swapped in or deepfake audio being used in phishing and identity theft scams. 

For the famous, these deepfakes can seriously damage someone’s reputation and cause them to lose sponsors or project deals. For the average person, deepfaking can cause people to lose their jobs or be embarrassed in front of their community.

Vulnerability to counterattacks

A deepfake doesn't just hurt your reputation; it can also make it seem like you've done something illegal or offended someone on purpose. You might get investigated for something you didn’t do.

How to spot deepfakes: 4 steps

As deepfakes proliferate on the internet, companies like Microsoft are providing watermark credentials to fight misinformation. Meta, for example, is holding deepfake detection challenges to teach its machine learning algorithms how to spot them. Likewise, cybersecurity companies like Deeptrace are being founded specifically to fight the scourge of deepfakes.

The problem: Deepfake technology keeps getting better, making it harder to tell what's real. Here are four signs.

Step 1: Check if the words or sounds match up with the video

AI-generated deepfakes can’t always match up the lip movements exactly with the audio. So, if you’re watching a video and feel like the sound is off a bit, it might be a deepfake. 

Take this deepfake video created by the Israeli startup Canny AI. It used footage of a speech Mark Zuckerberg gave in 2017 to make it seem like he was saying Facebook steals and controls people’s data. Though it’s a pretty convincing deepfake, as you watch, you can see the words and audio are a little off—signaling this is indeed fake.

Step 2: Check the source

An easy way to check if something may be a deepfake is the source of the material. For example, although this TikTok user has a very convincing Tom Cruise deepfake, his handle is @deeptomcruise. He’s not trying to hide that his entire channel is full of deepfake videos of Tom Cruise—meant to entertain.

Step 3: Check the visuals

The visual aspects of a video can tell you a lot. Check the lighting, the body posture, the background, body and head positioning of the subject to spot whether it’s a deepfake. 

There were deepfake videos in March 2022 showing Ukrainian President Zelensky saying he wanted Ukranians to surrender to Russian soldiers. A lot of Ukrainians laughed at the badly constructed videos and realized they were fakes right away because:

• Zelensky’s head and neck were two different colors with two different light sources
• His head position never changed and looked a bit unnatural
• His body never moved
• The image was blurry overall
• There were points where it looked as though his face would tick or glitch

Step 4: Check the audio

Most people focus on the visual aspects of a deepfake, but sometimes, the audio is where you’ll find the clues. If the sound is choppy, the audio levels change, or it feels like even the voice itself isn’t quite right—those are big clues that you may be watching a deepfake.

Pay special attention to any breath sounds or lack of breath sounds if the video shows the speaker breathing heavily or sighing. For example, although Morgan Freeman's voice in this video is similar to the actor’s voice, some words and phrases don’t sound exactly right. 

What are deepfakes FAQs

Are deepfakes illegal?

Deepfake content is technically legal—so long as you have permission of the person you’re imitating. If not, you can sue in a civil court for having your likeness used without consent, especially if you can prove the use of deepfakes has traumatized you or damaged your reputation. 

What were deepfakes made for?

Deepfakes were initially created to synthesize new facial expressions using audio content in a program called Video Rewrite Program. It took the audio of a video and modeled the lips and the person’s face in a 3D space so the video portion would match (lip-syncing) the audio content.

How do people create deepfakes?

Deepfake videos are created by using a number of technologies, including deep learning, facial recognition algorithms, and techniques like variational auto-encoders (VAE) and generative adversarial networks (GANs). The AI trains the encoders to the specific face or faces you wish to swap. It can create face content using that training data.