91��ɫ

Researchers closer to understanding how brains and AI differ in identifying facial expressions

The research is important for building more brain-like AI for understanding social perception that could in the future explain differences in autism, for example, leading to improved clinical tools

TORONTO, Aug. 27, 2026 – The human brain can recognize a person’s face and their expression in less than a blink of an eye, but current artificial intelligence (AI) models, although fairly accurate at doing the same, use a different process than a brain and that may not be good enough for applications in health care or education, according to new research out of 91��ɫ.

“Faces are critical to human communication. 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. We need to understand how those AI systems arrive at their judgments and where their interpretation of human social signals differs from ours,” says 91��ɫ senior author and Assistant Professor Kohitij Kar. “A major long-term benefit is that this work gives us a way to move beyond simply describing differences in social perception and toward understanding the neural mechanisms that produce them.”

The study, , published in Nature Communications.

A first step toward that goal is understanding how the brain represents both who a person is and what facial expression they are making, and then asking whether computational models capture those same representations.

“We can measure which individual facial expressions humans find easy or difficult to interpret, determine whether the same patterns are present in a carefully validated animal model, measure the underlying neural activity directly, and then use AI models to turn those biological observations into testable computational explanations,” says Kar, of the Faculty of Science and a member of the 91��ɫ-led Connected Minds.

For this study, postdoctoral researcher and lead author Maren Wehrheim, and team, tested and compared 11 artificial neural network (ANN) models, 290 humans and two non-human primates using 360 images of 12 people’s faces expressing anger, disgust, fear, joy, sadness and shame at five different strengths.

The researchers first asked whether non-human primates showed structured patterns of facial-expression perception similar to those seen in humans. They then tested whether ANNs could reproduce not only overall accuracy, but also the characteristic image-by-image judgments of the primates and the neural representations recorded from the monkeys’ brains. This allowed the team to test and constrain computational explanations of facial-expression perception.

An image of a face making one of the six expressions was shown for 200 milliseconds to keep the behavioral response limited to perceptual processing only. The human participants and non-human primates were then shown two additional images of the same person, to keep the focus on facial expression and not on their identity, from which they had to pick the correct facial expression.

“People became more accurate as expressions grew stronger, but some faces were consistently easier to judge than others. These image-by-image patterns created a fingerprint of human perception,” says Kar.

The non-human primates were able to identify different expressions, displaying similar structured, reproducible behavioral patterns, and although not as accurate as humans, they had an ability to perform the task with some differences. Both found the same faces difficult to judge and when it came to identifying shame, and less so anger and disgust, non-human primates were almost as accurate as humans.

As for ANNs, many were highly accurate in choosing the correct facial expression, however, the researchers were surprised that broadly trained object-recognition models did better at replicating the monkeys’ characteristic errors and behaviour than those specifically trained on faces or facial muscle movements.

“The finding highlights an important lesson for AI. 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,” says Kar, the Canada Research Chair in Visual Neuroscience and a member of 91��ɫ’s Centre for Vision Research and Centre for Integrative and Applied Neuroscience.

The researchers then went beyond behaviour and AI by recording activity from 308 sites in macaque inferior temporal (IT) cortex as the animals viewed the faces. IT is a high-level visual area important for recognizing objects and faces. By focusing on this visual pathway, the study examined the perceptual computations supporting expression discrimination rather than higher-order emotional or affective responses.

They found it matched the monkeys’ behaviour better than the ANNs. “Surprisingly, the strongest match to behavior appeared around 70 to 100 milliseconds after the image appeared, while later activity, although better at classifying the expression labels, was less similar,” says Kar.

This finding points to a heterogeneous coding architecture where facial identity processing and emotional expression overlap in the temporal cortex of the brain, indicating that the brain may act on an early visual signal even as it continues to process and refine the image.

The study validates the non-human primates as an appropriate animal model for exploring human facial expression processing and establishes a foundation for similar research and the potential for it to translate to human social cognition.

“Although, the study looked at visual discrimination, not emotional understanding, it revealed how quickly the primate brain turns a face into a meaningful perception. The findings provide a benchmark for building more brain-like AI and a framework for studying differences in social perception,” says Kar.

The current study did not examine autism directly, but Kar says the framework now makes it possible to ask a more mechanistic question about social-perception differences. “For autism in particular, we ultimately want to move from saying that two people perceive the same facial expression differently to being able to explain what visual information was represented differently, what neural computation produced that difference, and whether a computational model can predict it. Differences in facial-expression perception are well documented in autism, but clinical measures often tell us that perception differs without revealing why. That distinction will ultimately matter clinically.”

Rather than treating social perception as a single ability that is atypical in a person, this approach could eventually make it possible to build individualized perceptual profiles that identify which aspects of visual processing differ for a particular person.

“In the longer term, such profiles could help develop more sensitive behavioral assessments, identify subgroups that may benefit from different types of support, and provide quantitative measures for evaluating whether an intervention is actually changing the perceptual process it is intended to affect,” says Kar.

“Over the longer term, the better those computational models become, the more they may allow us to replace exploratory biological experiments with simulations and reserve animal research for the smaller set of causal questions that genuinely cannot be answered in any other way.”

Next steps are to apply this framework directly to autism – – and continue to improve the AI models.