When people judge whether a face is made of artificial intelligence or real, they come up with it about half the time. We could have guessed just as easily.
shows one A study conducted by researchers Sophie J. Nightingale and Hani Farid.
They combined psychology and computer science to see if people can distinguish fake faces created by artificial intelligence (AI) from real ones — and how we deal with fake faces.
– Our assessment is that the faces created by artificial intelligence are not only realistic scary. The researchers, working at Lancaster University and the University of California, respectively, wrote that the synthesis engine is capable of creating faces that are indistinguishable from real faces and that appear more reliable than real ones.
Generator and special
Behind the experiment is what is called a generative adversarial network (GAN). This technology is popular for superimposing content such as images, video, or audio.
A GAN puts two different neural networks against each other, one generator and one differentiator. The generator is trained to produce images, and the highlighter is trained to distinguish fake and real images from each other.
The generator starts with a random series of pixels, and through many iterations (iterations) it learns to make a fake but realistic face. For each iteration, the highlighter receives the images, along with the real ones, and separates the fake images provided by the generator. The highlighter “punches” the fake generator, and the generator learns to synthesize increasingly realistic faces.
This continues until the discriminator is no longer able to distinguish between false and true.
Researchers believe that despite advances in developing automated technologies to detect AI-generated content, often called deepfakes, the technologies we have today are neither effective nor accurate enough to handle the daily influx of downloads of such content. .
Users of web solutions must distinguish the real from the fake. Therefore, we conducted a series of observational studies to see if participants could distinguish between GAN-synthesized fresh faces from real faces and what level of confidence the faces elicited, they wrote in the article.
guess right half the time
In the first experiment, 315 participants looked through 800 pictures of faces, 128 of which were fake. The task was to distinguish the real from the fake.
The mean score was 48.2%. In other words, participants can easily guess blindly. The chance of getting it right is 50 percent.
But in the next experiment, the researchers trained a new group of people what to look for in order to spot the false images.
However, the average score rose to only 59 percent. Participants also received feedback along the way, but it turned out to have little effect.
Participants were the worst at distinguishing fake from real when there were pictures of white men. The researchers believe that the overrepresentation of white men in algorithm training made models more adept at fitting such images accurately.
In a third experiment, 223 people had to rate the reliability of the people in the pictures.
Trustworthy
Faces contain a lot of information, and we spend only a few milliseconds drawing conclusions about personality traits, such as how reliable a person is, the researchers wrote.
The hypothesis was that a gut feeling could help us distinguish between fake and real faces.
But the answer is not encouraging, on the contrary.
Participants had to rate the images on a scale of one to seven. The average score on the synthetic images was 4.82 versus 4.48 on the real images. The researchers believe that although the difference is just under 8 percentage points, it is significant.
There is little difference in how faces of different skin tones are evaluated, but women are seen as more trustworthy than men. The researchers wrote that smiling faces are largely reliable, but that a large portion of fake and real photos are of people smiling, so this alone cannot explain the difference.
They still have a theory.
Compound faces tend to look more average, which in itself increases our confidence in the people behind us.
Do the risks outweigh the benefits?
Although the development can be seen as a success for the technologists behind it, it is not without challenges.
The most convincing fake profiles on the internet, synthetic audio and video used for scams and disinformation, and fake videos that hang around or humiliate individuals. Researchers believe that all of this can have serious consequences for individuals, society and democracy.
– Therefore, we encourage those who work to develop this type of technology, to think about whether the risks outweigh the benefits, they write and advocate for both technological guidelines and solutions, such as watermarks that reveal false content.
– Since one of the biggest threats is the very easy access to this type of technology in the end, we recommend that you think twice before that without criticism and without restrictions, share the code publicly so that anyone can use it in any favorite application, the researchers wrote.
And if you’re wondering, this photo shows pairs of similar faces, real on the left and synthetic on the right.
“Web specialist. Lifelong zombie maven. Coffee ninja. Hipster-friendly analyst.”
