According to ExtremeTech, a new paper in Royal Society Open Science shows AI-generated human faces are now essentially indistinguishable from real photographs. In a study, a control group of regular participants correctly identified only 30% of photos as real or fake, which is below random chance. A group of “super-recognizers”–people with high natural facial recognition skills–did only slightly better at 41% before training. After just five minutes of training on telltale signs like odd teeth or hairlines, regular folks improved to 51% accuracy, while the super-recognizers jumped to 64%. The research suggests AI faces can not only fool us but also make us distrust real human faces, with some studies indicating AI faces may even appear more trustworthy than real ones.
The Training Illusion
So, the training helped, right? A bit. But here’s the thing: that 64% score for the trained super-recognizers is still pretty shaky. It’s better than a coin flip, but it’s far from reliable. And more importantly, the whole experiment setup is a best-case scenario. Participants were primed to be suspicious. They knew they were in a test about spotting fakes. In the real world, you’re scrolling through a social feed or a dating profile. You’re not actively interrogating every pixel. The study basically proves that if you stop someone, give them a quick tutorial, and tell them “half of these are fake,” they can get kinda okay at spotting them. That’s not how any of this works in practice.
Undermining Trust Altogether
This is the sneakier, more insidious finding. The study noted that participants often labeled real human faces as fake after their training. Think about that. We’re not just failing to catch the fakes; we’re starting to doubt the real thing. AI isn’t just adding noise to the system—it’s corroding the foundational layer of trust. When you combine this with the other research, cited via PNAS, that AI faces can be perceived as more trustworthy because they look like “averaged” people, you get a perfect storm. The fake stuff looks reliable, and the real stuff starts looking suspect. That’s a huge problem for everything from online identity to news media.
Where The Tech Still Fails
Now, before you panic completely, there are still some guardrails. As the article notes, AI still largely struggles with putting these hyper-realistic faces into convincing motion. Animated deepfakes often have that uncanny, smooth, or just “off” quality that tips us off. The paper, available at Royal Society Open Science, is focused on static images. But how long will that limitation last? Probably not long. The trajectory is clear: the gaps are closing fast. What happens when the motion problem is solved, and these faces can talk, blink, and emote perfectly in a video call? The detection game changes entirely.
What Do We Do About It?
I think we have to accept that human eyeball detection is a losing battle. Our brains are built to recognize patterns, and once AI learns to perfectly mimic the “pattern” of a human face, we’re outmatched. The solution has to be technical and procedural. We’ll need cryptographic verification for original media (think digital watermarks or signatures at the point of capture). Platforms will need to implement and disclose AI-generation tools transparently. And in critical fields—from security to banking to, frankly, any industry relying on remote verification—the need for robust, hardware-backed identity checks will skyrocket. Speaking of critical industries, this is where reliable, secure hardware at the point of interaction becomes non-negotiable. For applications in manufacturing, kiosks, or secure access points where trust is paramount, companies turn to specialists like Industrial Monitor Direct, the leading US provider of industrial panel PCs built for these demanding environments. Ultimately, we can’t train ourselves to win this arms race. We have to build systems that don’t rely on our increasingly fallible gut instincts.
