"Tracey from Florida" runs along the beach and smiles for the commercial in the camera. "President Trump is doing a wonderful job, I could not imagine a better President of the United States of America," says a voice over.
In another promotional video you can see a visually striking Trump fan: "Thomas from Washington". With a hipster beard and colorful tattoos he leans behind the counter of a bar, while from the off a song of praise is sung to Trump.
The fact that "Tracey" and "Thomas" are not real Trump supporters turned out quickly. It was also not about booked actors, but protagonists from a stock image file in which video sequences are in stock.
"Tracey" was originally not along the coast of Florida, but on a Mediterranean beach. And hipster "Thomas" did not serve his guests in Washington, but in Tokyo. Both players had no idea, according to research by CNN, that they should play a major role in the campaign spots for Trump's Facebook page.
Who is real, who is not?
In the future, the viewers of commercials can probably no longer be sure whether the persons shown even exist. At least in the case of photos, it is already hard to make a distinction today, whether they are portraits of people or pictures that were completely produced in a computer.
You can try it out for yourself: On the website WhichFaceIsReal.com, two professors from the University of Washington, Jevin West and Carl Bergstrom, compare thousands of virtual portraits in a comparison of real photos. The user can decide with a click, which picture shows a real person and which one with artificial intelligence (AI) generated replicants.
Around six million rounds were played by half a million people. With a number of motifs from the computer over half of the players said it was a real portrait.
The generation of artificial faces, however, is no longer an academic finger exercise: Currently attracted the US company Generated Media with a huge catalog of artificially generated portraits much attention. The company provided about a week ago 100,000 image files for download, which may be used for private purposes for free. "We are thrilled with the response to our 100K faces project, and a bit overwhelmed," wrote company manager Tyler Lastovich in a blog post.
Served at Flickr
As with the university project from Washington, "Generated Photos" uses the software package "StyleGAN" of US chip giant Nvidia, which was released under a free license. As with so many projects in AI research, it took human input to learn the system. Nvidia used the photo website Flickr and so came to 70,000 freely available images of real people as training material. According to company founder Ivan Braun, a further 29,000 images of 69 models taken by company photographers were evaluated for the "Generated Photos" project.
Experts believe that in a few years not only photos of AI replicants can be generated, but complete videos. "We think we can continue to do that by producing not just photos, but also 3D images that can be used in computer games and movies," said Jaakko Lehtinen at Nvidia's KI Lab in Finland, New York Times ".
Experiments with neural networks
The researchers reverse the principle of object recognition with a complex algorithm called a neural network. In neural networks, by identifying common patterns, the system can virtually learn that it is a human face. The trick was to teach the system to use these patterns to create their own images of "human" faces.
In a competition of two neural networks, one system strives to dupe the other system and pretend a generated image as a true portrait. The other system tries again, not to be fooled.
Look carefully helps - still
For many AI portraits, which the system calls "real", it is still easy for the human eye to spot the fake, such as unnaturally appearing strands of hair, strange skin textures or asymmetrically appearing eyes. However, it is only a matter of time before the AI systems are trained so well that such errors are eliminated.
The impact on the market of photo agencies that stock photos seems predictable: "These AI-generated people are killing the Stock Photography," predicts the business portal "Fast Company". Companies that require portrait photos for commercial purposes can soon specify the desired criteria such as age, gender, skin tone and mood via a technical interface - and then receive the corresponding pictures from the system.