Deepfakes: Understanding the Basics

In today’s episode, we'll explore four common types of deepfakes and gain a basic understanding of this  use of artificial intelligence (AI). Visit AVID Open Access to learn more.

Transcript

#271 — Deepfakes: Understanding the Basics

12 min
AVID Open Access

Keywords

images, ai, voice, video, technology, tool, type, create, speak, fake, morgan freeman, generated, clip, face, develop, number, text, convincing, manipulated, translated

Speakers

Paul (87%), AI (7%), Transition (2%), AI (2%)

Paul Beckermann  0:01  

Welcome to Tech Talk for Teachers. I'm your host, Paul Beckermann. 

Transition Music  0:06  

Check it out. Check it out. Check it out. What's in the toolkit? What is in the toolkit? So, what's in the toolkit? Check it out. 

Paul Beckermann  0:16  

The topic of today's episode is Deepfakes: Understanding the Basics. It's very likely that you've heard the term deepfake being used lately. It's been in the news and it's a common topic during conversations about generative artificial intelligence. But what exactly is a deepfake? Let's look at a few definitions. The Oxford English Dictionary defines a deepfake as "Any of various media, especially a video, that has been digitally manipulated to replace one person's likeness convincingly with that of another, often used maliciously to show someone doing something that he or she did not do." Here's another definition, this one from Merriam-Webster Dictionary, it defines a deepfake as "An image or recording that has been convincingly altered and manipulated to misrepresent someone is doing or saying something that was not actually done or said." And the Government Accountability Office defines a deepfake as "A video, photo, or audio recording that seems real, but has been manipulated with AI." So essentially, a deepfake is some sort of media, a video, image, or audio recording that sounds or looks real, but is actually fake. This type of transformative technology is poised to have a significant impact on society. It's already been used to produce fake news stories, to create clickbait and to mislead voters. In a democracy where citizens must be able to make informed voting decisions, it's imperative that we all become digitally literate in this new technology. That means adults and students alike. The first step in developing deepfake literacy is understanding the basics, and in today's episode, I'm going to share with you some of those basics. Let's start by looking at four common types of deepfakes. 

Transition Music  2:12  

Let's count it, let's count it. Let's count it down. 

Paul Beckermann  2:16  

Number one, face swapping. This is exactly what it sounds like. One person's face is swapped out or replaced by another face and a video or a picture. Some of these are created by seasoned pros with sophisticated software, and some by average consumers who have downloaded apps like Apple's Face Swap AI photo app. People use the app to put their faces on celebrity images or to see what they look like with someone else's hairstyle. A lot of times people will use these apps and techniques simply to have fun. The YouTube channel Ctrl Shift Face is a collection of video face swaps meant to entertain. There's a very convincing one that has Jack Nicholson's face swapped out for Jim Carrey's in a famous scene from The Shining. Honestly, if I didn't know it was a face swap, I'm not sure I'd know that wasn't the original footage from the film. Of course, I do know that Jim Carrey wasn't in the film, so I wasn't fooled. But if an audience wasn't familiar with the original context, they could easily be fooled by this deepfake face swap. Now a deepfake of a movie clip like this probably won't cause any major harm. It's mostly for fun. But consider if a face were swapped out in a photo or a video of a crime scene or at a political protest. In these contexts, the misleading image could easily damage your reputation, or lead to some real life negative consequences. 

Number two, face manipulation. In this case, one person's face is not replaced by another's, instead of video of someone is manipulated to make it look like they're saying something that they never said. For example, there's a famous YouTube video called This Is Not Morgan Freeman, that shows an AI deep fake of the actor demonstrating how real it can appear. It starts with the face of the actor on the screen saying "I am not Morgan Freeman, and what you see is not real." While the lips seem a little off to me, it really does sound like Morgan Freeman. And if I weren't paying close attention, I could easily be convinced that it was indeed Morgan Freeman who was speaking. There's been increasing concern that this type of deepfake technology might be used to spread election misinformation. A candidate for office might be made to appear to be saying something that they never said. Because a video is showing them saying it. It can be very convincing, and viewers may believe that it's real. 

Number three, voice synthesis. This could mean a couple things. On one hand, it might mean converting text to speech, like with the Chrome extension Speechify. This can be a really great accessibility tool for people who struggle to read. It can also streamline commercial production projects. My son's company has used this type of technology to produce AI generated voiceovers for training videos. Hollywood recently made famous use of voice synthesis in the movie Top Gun. Val Kilmer had lost his voice due to throat cancer, and this technology was used to allow him to speak in the movie. He wasn't really speaking, but the synthetic version of his voice made it sound like he was, and it was convincing. Well, these are positive legitimate uses for this type of AI, it's also been used for nefarious purposes as well. Voice synthesis has been used to make it sound like someone is saying something that they never actually said. In fact, this happened recently during the New Hampshire primary elections when a fake robo call went out using President Biden's voice to discourage voters from going to the polls. 

Number four, AI generated images. Within the last year, dozens of AI image generations have been released, and many of them are free. These includes tool like DALL-E, Midjourney, and Adobe Firefly. Users essentially type in a description as a text prompt, and the tool produces AI generated images. These pictures are created based on millions of images that have been scanned in order to train the AI generator how to construct a new image. This has caused concern on multiple levels. Many photographers and artists who have images online believe their work has been used without their consent, and that these AI companies are essentially stealing their style and techniques to create these new images. To offset this concern, some companies like Adobe are only using images for which they own all the rights. Recently, deepfake images of Taylor Swift went viral on the social media platform X, being viewed 47 million times before X took them down. There been other examples of realistic looking images that were deepfakes, such as the image showing Donald Trump getting tackled and arrested by New York City police officers. Or another one that appeared to show Vladimir Putin locked up in prison. Again, it never happened. Both of these images were fake. But they looked real enough that they got circulated on social media with little or no context. And some people believed that they were real. The proliferation of these four types of deepfakes is growing quickly. A 2023 Reuters article about deepfakes sites that from May 22, to May 2023, deepfake videos tripled. It also reports that there were eight times as many voice deepfakes and 2023 as there were the previous year. It's not surprising that instances of deepfakes are on the rise. The tools to create this fake media are becoming more and more commonplace, and many are available either for free or a small fee. To give you an idea of how convincing these tools are, I put together a few examples using two free video tools, Speechify and ElevenLabs. These tools have premium subscriptions as well, but I'll just use the free features.

Transition Music with Rena's Children  8:12  

How do I use this? Integration inspiration, integration ideas! 

Paul Beckermann  8:16  

Let me start with Speechify. For this one, I created my free account and uploaded a three minute clip of me speaking from an earlier episode of Tech Talk for Teachers Speechify took this clip to learn my voice and how I speak and then generated this clip: 

AI Paul  8:18  

 Speechify has revolutionized my productivity with its custom text to speech voices. Partnering with Speechify allowed me to clone my own voice, making it faster than ever to create new audio and video content.

Paul Beckermann  8:48  

That was the AI talking. I never said those words in the clip I uploaded. In less than two minutes, I was able to clone my own voice and generate audio that I never actually spoke. The next tool I tried out was ElevenLabs. This is the tool that was used to create the phony Biden robo call that I mentioned earlier. You can't do that voice cloning on this free version, but it only costs $1 a month to be able to use that feature. For my test, I tried out their text to speech tool and their translation tool. For the text to speech conversion. I used one of the built in voices because I didn't have the premium version. I chose the voice of Sarah, pasted in some text and hit generate. Here's what I got: 

AI Sarah  9:31  

One of the areas of artificial intelligence that is getting a lot of attention is image generation. In the simplest terms, you type in a text description of the picture you'd like generated and click enter.

Paul Beckermann  9:45  

I think that's really good quality. And because it's not trying to deceptively clone a real speaker's voice, I can see genuine and honest uses for that technology. For my next test, I uploaded a video of myself saying a couple sentences Here's the original audio. "My name is Paul Beckermann and I'm testing translation software. I hope it works well because this could be a really helpful tool for students and families who do not speak English as their primary language." Then I had ElevenLabs translate that clip into Spanish they call this dubbing here's that audio. 

AI Paul  10:22  

Mi nombre es Paul Beckermann y estoy probando el software de traduccion. Espero que funcione bien por que esta pordria ser una herramienta realmente útil para estudiantes y familias que no hablan inglés como idioma principal. 

Paul Beckermann  10:35  

And then I had it translated to French. 

AI Paul  10:38  

Je m'appelle Paul Beckermann et je teste un logiciel de traduction. J'espère que cela fonctionnera bien car pourrait être un outil très utile pour les étudiants et les familles qui ne parlent pas l'anglais comme langue principale. 

Paul Beckermann  10:52  

To me, that's really impressive. I even played it for my friend who speaks those languages. And she said it was very convincing. Now I should mention that I also got a video version of those translations and I didn't find it nearly as convincing. My mouth didn't really line up well with the translated pronunciations, and it didn't look real. However, if you were needing to communicate with people in other languages, it could be powerful. Overall, this technology is impressive, powerful, and potentially a little worrisome, because this technology is not going away. In fact, I think it's going to continue to get more sophisticated, we need to become savvy digital consumers. The first step for all of us in doing that is to know that this technology exists, and have a basic awareness of how it works. I hope this episode of Tech Talk for Teachers has helped to give you that foundation. The second step is to develop media literacy around this deep fake technology, so we can protect ourselves from being misled. It'll be essential that we help our students develop these skills as well. Tune in next week as I continue this conversation and dive into strategies for identifying deepfakes and developing that literacy. 

To learn more about today's topic and explore other free resources, visit AVIDOpenAccess.org. And of course, be sure to join Rena, Winston, and me every Wednesday for our full-length podcast, Unpacking Education, where we're joined by exceptional guests and explore education topics that are important to you. Thanks for listening. Take care, and thanks for all you do. You make a difference.