The human brain can read a face in less time than it takes to blink — yet even the smartest computer models still can't quite do it the way we do. That gap matters, say researchers at York University in Toronto, and it could decide whether artificial intelligence is truly ready for jobs in health care and classrooms.
"Faces are critical to human communication," says Kohitij Kar, the study's senior author and an assistant professor at York. "Very small changes in facial appearance can influence how we understand a conversation, whether we think someone is comfortable or distressed, and how we respond socially." If AI reads those signals differently than we do, he argues, we need to know exactly where they disagree.
To find out, postdoctoral researcher Maren Wehrheim and her team staged a giant face-reading experiment. They showed 290 humans, two monkeys, and eleven artificial neural networks — computer programs loosely inspired by the brain — a set of 360 images of 12 people's faces. The faces showed six expressions — anger, disgust, fear, joy, sadness, and shame — at five different levels of strength. Each image flashed for just 200 milliseconds, fast enough to force the viewers to rely purely on perception, not careful thought.
The results, published in the journal Nature Communications, produced a kind of fingerprint of how living brains read emotions. People got more accurate as expressions grew stronger, but some faces were always trickier to judge than others. The monkeys showed the same structured patterns, struggling with the very same faces as the humans. When it came to spotting shame — and, to a lesser degree, anger and disgust — the monkeys were nearly as accurate as people.
Then came the surprise. Many of the AI models nailed the task, choosing the right expression again and again. But the models that best copied the monkeys' specific mistakes weren't the ones specially trained on faces at all. Broadly trained object-recognition models did a better job of recreating how the monkeys actually thought, including their characteristic errors.
That finding flips a common assumption on its head, Kar says. "Getting the right answer is not the same as solving the problem in a brain-like way. A specialized system may become efficient while discarding information that still shapes biological perception."
In other words, an AI can score perfect on a test while missing the subtle social cues that humans read without thinking. That matters because machines trained to detect mental distress or to tutor students would be working with our most human signals. The team's long-term goal is to use these brain recordings to build a bridge between biology and machine intelligence — turning what we observe in humans and monkeys into computational explanations they can actually test. It's a step toward AI that doesn't just answer correctly, but understands people the way people understand each other.