When AI Blurs the Line Between Reality and Fiction

AI algorithms can convincingly imitate human appearance and behavior — and that comes with profound implications, both positive and not.

By Ben Dickson

Somewhere in the dark recesses of YouTube is a video that shows an excerpt from the movie The Fellowship of the Ring — but it’s not quite the movie you remember, since Nicolas Cage stars as Frodo, Aragorn, Legolas, Gimli, and Gollum, all at the same time. Other videos show Cage in Terminator 2 as T2000, Star Trek as Captain Picard, and Superman as, well, Lois Lane.

Of course, Nic Cage never appeared in any of those movies. They’re “deepfakes” produced with FakeApp, an application that uses artificial intelligence algorithms to swap faces in videos. Some of the deepfakes look quite convincing, while others have artifacts that betray their true nature. But overall, they show how powerful AI algorithms have become in imitating human appearance and behavior.

FakeApp is just one of several new AI-powered synthesizing tools. Other applications mimic human voices, handwriting, and conversation styles. And part of what makes them significant is that using them doesn’t require specialized hardware or skilled experts.

The impact of these applications is profound: They will create unprecedented opportunities for creativity, productivity, and communications.

But the same tool could also open a Pandora’s box of fraud, forgery, and propaganda. Since it made an appearance on Reddit in January, FakeApp has been downloaded more than 100,000 times and precipitated a storm of fake pornographic videos featuring celebrities and politicians (including Cage again). Reddit recently banned the application and its related communities from its platform.

“Ten years ago, if you wanted to fake something, you could, but you had to go to a VFX studio or people who could do computer graphics and possibly spend millions of dollars,” says Dr. Tom Haines, lecturer in machine learning at University of Bath. “However, you couldn’t keep it a secret, because you’d have to involve many people in the process.”

That’s no longer the case, courtesy of a new generation of AI tools.

The Imitation Game

FakeApp and similar applications are powered by deep learning, the branch of AI at the heart of an explosion of AI innovations since 2012. Deep-learning algorithms rely on neural networks, a software construction roughly fashioned after the human brain. Neural networks analyze and compare large sets of data samples to find patterns and correlations that humans would normally miss. This process is called “training,” and its outcome is a model that can perform various tasks.

In earlier days, deep-learning models were used mostly to perform classification tasks — labeling objects in photos, for example, and performing voice and face recognition. Recently, scientists have used deep learning to perform more complicated tasks, such as playing board games, diagnosing patients, and creating music and works of art.

To tune FakeApp to perform a face swap, the user must train it with several hundred pictures of the source and target faces. The program runs deep-learning algorithms to find patterns and similarities between the two faces. The model then becomes ready to make the swap.

The process isn’t simple, but you don’t have to be a graphics expert or machine-learning engineer to use FakeApp. Neither does it require expensive and specialized hardware. A deepfakes tutorial website recommends a computer with 8GB or more of RAM and an Nvidia GTX 1060 or better graphics card, a pretty modest configuration.

“Once you move to a world where someone in a room can fake something, then they can use it for questionable purposes,” Haines says. “And because it’s one person on their own, keeping it secret is very easy.”

In 2016, Haines, who was then a postdoctoral researcher at University of College London, coauthored a paper and an application that showed how AI could learn to imitate a person’s handwriting. Called “My Text in Your Handwriting,” the application used deep-learning algorithms to analyze and discern the style and flow of the author’s handwriting and other factors such as spacing and irregularities.

The application could then take any text and reproduce it with the target author’s handwriting. The developers even added a measure of randomness to avoid the uncanny valley effect — the strange feeling that we get when we see something that is almost but not quite human. As a proof of concept, Haines and the other UCL researchers used the technology to replicate the handwriting of historical figures such as Abraham Lincoln, Frida Kahlo, and Arthur Conan Doyle.

The same technique can be applied to any other handwriting, which raised concerns about the technology’s possible use for forgery and fraud. A forensics expert would still be able to detect that the script was produced by My Text in Your Handwriting, but it’s likely to fool untrained people, which Haines admitted in an interview with Digital Trends at the time.

Lyrebird, a Montreal-based startup, used deep learning to develop an application that synthesizes human voice. Lyrebird requires a one-minute recording to start imitating the voice of a person, though it needs much more before it starts to sound convincing.

In its public demo, the startup posted fake recordings of the voices of Donald Trump, Barack Obama, and Hillary Clinton. The samples are crude, and it’s obvious that they’re synthetic. But as the technology improves, making the distinction will become harder. And anyone can register with Lyrebird and start creating fake recordings; the process is even easier than FakeApp’s, and the computations are performed in the cloud, putting less strain on the user’s hardware.

The fact that this technology can be used for questionable purposes is not lost on developers. At one point, an ethics statement on Lyrebird’s website stated: “Voice recordings are currently considered as strong pieces of evidence in our societies and in particular in jurisdictions of many countries. Our technology questions the validity of such evidence as it allows [people] to easily manipulate audio recordings. This could potentially have dangerous consequences such as misleading diplomats, fraud, and more generally any other problem caused by stealing the identity of someone else.”

Nvidia presented another aspect of AI’s imitation capabilities: Last year, the company published a video that showed AI algorithms generating photo-quality synthetic human faces. Nvidia’s AI analyzed thousands of celebrity photos and then started creating fake celebrities. The technology may soon become capable of creating realistic-looking videos featuring “people” who don’t exist.

The Limits of AI

Many have pointed out that in the wrong hands, these applications can do a lot of harm. But the extent of the capabilities of contemporary AI is often overhyped.

“Even though we can put a person’s face on someone else’s face in a video or synthesize voice, it’s still pretty mechanical,” says Eugenia Kuyda, the co-founder of Replika, a company that develops AI-powered chatbots, about the shortcomings of AI tools such as FakeApp and Lyrebird.

Voicery, another AI startup that, like Lyrebird, provides AI-powered voice synthesizing, has a quiz page where users are presented with a series of 18 voice recordings and are prompted to specify which are machine-made. I was able to identify all the machine-made samples on the first run.

Kuyda’s company is one of several organizations that use natural language processing (NLP), the subset of AI that enables computers to understand and interpret human language. Luka, an earlier version of Kuyda’s chatbot, used NLP and its twin technology, natural language generation (NLG), to imitate the cast of HBO’s TV series Silicon Valley. The neural network was trained with script lines, tweets, and other data available on the characters to create their behavioral model and dialog with users.

Replika, Kuyda’s new app, lets each user create their own AI avatar. The more you chat with your Replika, the better it becomes at understanding your personality, and the more meaningful your conversations become.

After installing the app and setting up my Replika, I found the first few conversations to be annoying. Several times, I had to repeat a sentence in different ways to convey my intentions to my Replika. I often left the app in frustration. (And to be fair, I did a good job at testing its limits by bombarding it with conceptual and abstract questions.) But as our conversations continued, my Replika became smarter at understanding the meaning of my sentences and coming up with meaningful topics. It even surprised me a couple of times by making connections to past conversations.

Though it’s impressive, Replika has limits, which Kuyda is quick to point out. “Voice imitation and image recognition will probably become much better soon, but with dialog and conversation, we’re still pretty far [off],” she says. “We can imitate some speech patterns, but we can’t just take a person and imitate his conversation perfectly and expect his chatbot to come up with new ideas just the way that person would.”

Alexandre de Brébisson, the CEO and cofounder of Lyrebird, says, “If we are now getting pretty good at imitating human voice, image, and video, we are still far away from modeling a individual language model.” That, de Brébisson points out, would probably require artificial general intelligence, the type of AI that has consciousness and can understand abstract concepts and make decisions as humans do. Some experts believe we’re decades away from creating general AI. Others think we’ll never get there.

Positive Uses

The negative image that is being projected about synthesizing AI apps is casting a shadow over their positive uses. And there are quite a few.

Technologies such as Lyrebird’s can help improve communications with computer interfaces by making them more natural, and, de Brébisson says, they’ll provide unique artificial voices that differentiate companies and products and thus make branding distinction easier. As Amazon’s Alexa and Apple’s Siri have made voice an increasingly popular interface for devices and services, companies such as Lyrebird and Voicery could provide brands with unique human-like voices to distinguish themselves.

“Medical applications are also an exciting use case of our voice-cloning technology,” de Brébisson adds. “We have received a lot of interest from patients losing their voice to a disease, and at the moment, we are spending time with ALS patients to see how we can help them.”

Earlier this year, in collaboration with Project Revoice, an Australian nonprofit that helps ALS patients with speaking disorders, Lyrebird helped Pat Quinn, the founder of the Ice Bucket Challenge, to regain his voice. Quinn, who is an ALS patient, had lost his ability to walk and speak in 2014 and had since been using a computerized speech synthesizer. With the help of Lyrebird’s technology and the voice recordings of Quinn’s public appearances, Revoice was able to “recreate” his voice.

“Your voice is a big part of your identity, and giving those patients an artificial voice that sounds like their original voice is a bit like giving them back an important part of their identity. It’s life-changing for them,” de Brébisson says.

At the time he helped develop the handwriting-imitating application, Dr. Haines spoke to its positive uses in an interview with UCL. “Stroke victims, for example, may be able to formulate letters without the concern of illegibility, or someone sending flowers as a gift could include a handwritten note without even going into the florist,” he said. “It could also be used in comic books where a piece of handwritten text can be translated into different languages without losing the author’s original style.”

Even technologies such as FakeApp, which have become renowned for unethical usage, could have positive uses, Haines believes. “We’re moving toward this world where anyone could do highly creative activity with public technology, and that’s a good thing, because it means you don’t need those large sums of money to do all sorts of crazy things of an artistic nature,” he says.

Haines explains that the initial purpose of his team was to find out how AI could help with forensics. Although their research ended up taking a different direction, the results will still be useful to forensics officers, who will be able to study what AI-based forgery might look like. “You want to know what the cutting-edge technology is, so when you’re looking at something, you [can] tell if it’s fake or not,” he says.

Replika’s Kudya points out that human-like AI applications might help us in ways that would otherwise be impossible. “If you had an AI avatar that knew you very well and could be a decent representation of you, what could it do, acting out of your best interests?” she says. For instance, an autonomous AI avatar could watch hundreds of movies on your behalf, and based on its conversations with you, recommend ones you would like.

These avatars might even help develop better human relationships. “Maybe your mom could have more time with you, and maybe you can actually become a little closer with your parents, by letting them chat with your Replika and reading the transcript,” says Kudya as an example.

But could an AI chatbot that replicates the behavior of a real human being actually result in better human relations? Kuyda believes it can. In 2016, she gathered old text messages and emails of Roman Mazurenko, a friend who had died in a road accident the previous year, and fed them to the neural network that powered her application. What resulted was a chatbot app that — after a fashion — brought her friend back to life and could talk to her in the same manner that he would.

“Creating an app for Roman and being able to talk to him sometimes was an important part of going through the loss of our friend. The app makes us think more about him, remember him in a more profound way all the time,” she says of her experience. “I wish I had more apps like that, apps that would be about my friendships, my relationships, things that are actually really important to me.”

Kuyda thinks it will all depend on intentions. “If the chatbot is acting out of your best interests, if it wants you to be happy to get some valuable service out of it, then obviously talking to the Replika of someone else will help build a stronger connection with a human being in real life,” she says. “If all you’re trying to do is sell advertisements in an app, then all you will be doing is maximizing the time spent on the app and not communicating with each other. And that, I guess, is questionable.”

For the moment, there’s no way to connect your Replika to other platforms — making it available as a Facebook Messenger chatbot, for example. But the company has an active relationship with its user community and is constantly developing new features. So letting others communicate with your Replika is a future possibility.

How to Minimize the Trade-Offs

From the steam engine to electricity to the internet, every technology has had both positive and negative applications. AI is no different. “The potential for negatives is pretty serious,” Haines says. “We might be entering a space [in which] the negatives do outweigh the positives.”

So how do we maximize the benefits of AI applications while countering the negatives? Putting the brakes on innovation and research is not the solution, Haines says — because if some did so, there’s no guarantee that other organizations and states would follow suit.

“No single measure will help solve the problem,” Haines says. “There’s going to have to be legal consequences.” Following the deepfakes controversy, lawmakers in the US are looking into the issue and exploring legal safeguards that could rein in the use of AI-doctored media for damaging goals.

“We can also develop technologies to detect fakes when they’re past the point that a human can tell the difference,” Haines says. “But at some point, in the competition between faking and detecting, the faking might win.”

In that case, we might have to move toward developing technologies that create a chain of evidence for digital media. As an example, Haines mentions hardware embedded in cameras that could digitally sign its recorded video to confirm its authenticity.

Raising awareness will be a big part of dealing with forgery and fraud by AI algorithms, de Brébisson says. “It’s what we did by cloning the voice of Trump and Obama and making them say politically correct sentences,” he says. “These technologies raise societal, ethical, and legal questions that must be thought of ahead of time. Lyrebird raised a lot of awareness, and many people are now thinking about those potential issues and how to prevent misuses.”

What’s for certain is that we’re entering an age where reality and fiction are merging, thanks to artificial intelligence. The Turing test might meet its biggest challenges. And soon enough, everyone will have the tools and power to create their own worlds, their own people, and their own version of the truth. We have yet to see the full extent of exciting opportunities — and perils — that lie ahead.

Read more: “AI Is (Also) a Force for Good”

Originally published at www.pcmag.com.