Dan Foy/Flickr

Without thinking about it, without even trying, we’ve created a surveillance apparatus more impressive than any new technology or any misreading of a law could dare to provide. We’ve given it life just by opening our eyes; exposing sensory organs so sharp in their ability to render reality that they override any other human sense just by default. Imagine being able to plug into the visual inputs of others, to capture and interpret what they see. You already have a sense of this idea from Blade Runner and subsequent decades spent rolling in the “enhance” trope.

The general concept—extracting information from eyeball reflections—is at the heart of a new study finding that humans are able to recognize faces reflected in the black pupils of eyeballs with about 80 percent accuracy. Remarkably, this recognition is possible using even kind of crappy images.

The study, led by the University of York’s Dr. Rob Jenkins, took passport style photographs of human faces, zoomed way the hell in on the eyeballs, and tasked participants with identifying faces reflected in the photo subject’s eyeball. They were able to identify correctly up to 71 percent of the time for faces they hadn’t been previously exposed to, and up to 84 percent of the time for faces they had encountered before.

“The pupil of the eye is like a black mirror,” Jenkins said in the study’s press notes. “To enhance the image, you have to zoom in and adjust the contrast. A face image that is recovered from a reflection in the subject’s eye is about 30,000 times smaller than the subject’s face.”

PLOS ONE/Dr Rob Jenkins

Think about that. You have the ability to recognize a face at 1/30,000 the size. The camera resolution used was 39 megapixels, a not terribly remarkable number. This is a task humans aren’t just good at, but excel at in a way bordering on strange.

“Obtaining optimal viewers [to interpret the images] may be more important than obtaining optimal images,” Jenkins added in an interview. “What I like about the study is that it stretches the limits of high resolution imaging, and stretches the limits of human vision, to the point where they overlap.”

Collecting these pupil reflections into any sort of imaging network is still a distant dream, but it’s certainly something that’s been considered. Call it sight mining. “As the environment becomes ever more populated with digital images, there is increasing interest in mining these images for hidden information,” Jenkins said. “Other research groups have described different methods for extracting hidden information from images. Ours is the first to demonstrate that eye reflections can be used to identify hidden bystanders.”

Surveillance aside, the near term implications for this kind of pupil reflection analysis are much more likely to involve busting child pornographers or crime scene witnesses. Photographers can no longer avoid being implicated in their own photos (at least of people).

Animated zoom on the cornea of a high-resolution photographic subject.

The York study circles back to larger problems in AI-based facial recognition. Ideally, we could skip the humans entirely, but this is the sort of thing that computers just can’t do very well.

“For example, automatic face recognition systems have tried to find facial signatures based on distances between facial features,” Jenkins said. “That is a dead end. Take the face images that we extracted from eye reflections in this study. Those images were only a few pixels high. You simply can’t pinpoint where facial features begin and end—there isn’t enough detail in the images. And yet humans who are familiar with those faces recognize them very well.”

“The only system that can reliably recognize poor quality face images is a human who is familiar with the faces concerned,” Jenkins continued. “We would do well to emulate that system.”