Years ago, I was having a conversation with someone – I don’t remember who – about dreams. I’ve always had dreams so lifelike and random that people think I’m making them up. Whoever this poor victim of my subconscious storytelling was, they told me something interesting: the people you see in your dreams are all people you’ve seen before. Whether they’re a celebrity, a family member, a close friend, or just someone you walked past on the street earlier that day – you’ve seen them before. The brain (according to whoever this was) is incapable of creating new faces. The neuroscience department at Stanford responded to this theory, essentially stating that it is “an impossible question to answer experimentally” (see https://neuroscience.stanford.edu/news/can-sleeping-brain-create-unique-people-waking-brain-has-never-seen), so we just don’t know, and there’s no way of knowing. It makes sense to me, though. If I’m writing a story, I can’t picture in my head an entirely new face for a character. I use myself or someone I know.
So, when I read Tom Simonite’s article in Wired about AI-generated faces, I was pretty spooked. If the theory about dreams is true, then this is something AI can do that human’s can’t. Needless to say, I won’t be mentioning this to my grandparents to feed their fear that robots are going to take over the world.
I played the “Which Face is Real” (http://www.whichfaceisreal.com/) game Simonite mentions. The program puts two faces, side by side, and the player clicks on the one he or she thinks is real.
I found some trends with the AI-generated faces: most of them weren’t blinking or looking in directions other than straight at the camera (oftentimes they appeared as having lazy eyes), but as we see in the second example, there are exceptions. To me, the image on the right looked creepily fake – but it was too obvious. The AI is smart. Simonite writes: “On average, players could identify the reals nearly 60 percent of the time on their first try. The bad news: Even with practice, their performance peaked at around 75 percent accuracy.” The fake faces could also be extremely high-quality images, as if they were taken by a professional camera, while some mirrored phone selfies. Sometimes it was obvious (proportions were off, backgrounds were general) and sometimes it was indiscernible.
I hope this doesn’t turn into a nightmare. I hope identity thieves and the like will just stay out of this. Maybe this technology is being safely guarded; I don’t know. What I do know is now I can go to whichfaceisreal.com and screenshot hundreds of fake faces – images I cannot even dream of. Where the internet’s quest for fake things will end, no one knows.
Simonite’s article: https://www.wired.com/story/artificial-intelligence-fake-fakes/