With AI able to replicate the human visage, skepticism is only natural. Let’s get through the basics and see how it actually works!
Artificial intelligence is now so good at generating human faces that most people can’t tell a real from a fake. The website This Person Does Not Exist, for instance, uses AI to generate human faces that are varying degrees of realistic—some scarily so.
Another called Generated Photos hosts a collection of AI-generated images, made using a private dataset of models, that can be used for any purpose including dating profiles.
And then, there’s Cambridge University’s Art AI project, which uses similar technology to create works of art that mirror visual hallmarks of famous artistic movements throughout history.
So, how does this technology work?
Image via Simone Golob.
To understand AI-generated faces, you need to understand GAN (generative adversarial network) technology, a machine-learning framework first developed back in 2014 by researcher Ian Goodfellow and his colleagues.
Let’s start by breaking down how GAN technology works to replicate art.
Think of the technology as having two competing models—the art forger and the art authority tasked with detecting fakes. The forger (generator) ingests a series of art references in order to learn how to create a painting. It learns all that goes into creating a work of art—composition, brushwork, subject matter, material, etc.
It then passes its forgery along to the art authority (discriminator), which has been ingesting a large volume of real paintings in order to develop a visual vocabulary.
GAN technology is a machine-learning framework. Image via Harry Haysom / Ikon Images.
If the discriminator determines an input to be genuine, its model is updated with this information, further reinforcing and developing its logic. Should the discriminator spot the fake, however, the forger’s model receives a message along the lines of:
“Hey, something in our method resulted in the product being detected as a fake. Let’s amend our methods so that next time we produce a painting, it’s more likely to fool the art authority.”
These cycles repeat millions of times, resulting in an overall more stable and consistent framework.
Images via Takahiro Suganuma and Rebecca Hendin / Ikon Images.
GAN technology works in a similar way to create AI-generated faces. There are two competing models. One ingests a large number of human references and learns what composes a human face. The other assembles disparate features into a wholly new and unique face. The fake faces are then fed back to the discriminator to determine whether they pass as real.
As unsettling as this process may sound, GAN machine learning isn’t the first man-made creation that’s influenced (or expanded) our understanding of what makes a human a human. There are many methods we use to enhance features we like and modify those we don’t, like plastic surgery.
With the right makeup and brushes, we can contour our faces to be almost unrecognizable. Or, we can pop on a TikTok filter to see what we looked like if we . . . looked nothing like ourselves.
Art has long been a challenger of self-image. Image via Simone Golob.
Our understanding of self has also been challenged by art for as long as we can remember—the exquisite corpse games of the Surrealists, video games like The Sims, and shows like Westworld. We’re endlessly asking if the sum of our parts is what makes us human, or if there’s something more.
The amount of open-ended philosophical questions relating to this subject of humanity is exactly why it’s our responsibility to ethically and comprehensively plan for the adaptation of such tech into our Shutterstock environment.
AI-generated content will be a constant check point in terms of ethics. Image via solarseven.
With the advent of AI-generated content, we are confronted with questions like:
How do we ensure this content is legally safe to license?How do we label this content on site in a way that doesn’t potentially alienate the contributor base at large?How do we apply existing filters, like age, gender, and race, to AI-generated content? Would this count as accurate and authentic representation?How does this AI-generated content hold up to curatorial ideas related to image authenticity? Do we have to redefine what it means for an image to be authentic?
Image via Simone Golob.
In searching our collection for a sample set of thought-provoking “machine-made” human content, I was hard-pressed to find images that adequately grasped the “excited but cautious” vibe.
So, I decided to search our collection to see how contributors have already been approaching explorations of humanity. Content that strives to answer the question: “What makes a human a human?”
Artificial intelligence is a hot topic. Take a look at a few of these, for instance:
How to Build the Perfect Thanksgiving Dinner, According to Data10 Tips for Choosing a Clickable Holiday Image, According to DataWhat Exactly Is a Deepfake and How Are They Made?Visual Language: Cybersecurity Awareness Through ImageryHow Artists Can Reach New Creative Heights Through Artificial Intelligence
Cover image via Artem Kovalenco.
The post AI-Generated Faces: What Makes a Human a Human? appeared first on The Shutterstock Blog.