KAIST develops AI that identifies autism by observing animal behavior — no prior training required

An AI that reads animal movements like words — and extracts meaning from them — has been developed by researchers in South Korea.

KAIST announced Tuesday that a research team led by Professor Kim Dae-su of the Department of Brain and Cognitive Sciences has developed an AI model called BehaVERT, which learns behavioral data as if it were language and independently identifies social behavior deficits in autism model mice.

The team built BehaVERT by converting skeletal movements of mice into "tokens" — units analogous to words in natural language — and training the model on that data. The model can independently discover core social behavior deficits in autism model mice without any prior domain knowledge.

The research presents a new AI approach to analyzing animal behavior as if it were language. By showing that AI can go beyond simple behavioral classification to understand the meaning of behavior and identify biologically significant cues, the work opens the door to a next-generation "Behavior Foundation Model" for drug development, psychiatric disorder research and behavioral genetics.

The team converted skeletal coordinates of body parts — including the nose, ears, spine, limbs and tail — into tokens and fed them into a BERT-based transformer model widely used in natural language processing. As a result, BehaVERT not only classifies behaviors but independently learns their meaning as they unfold over time.

The model outperformed existing state-of-the-art systems across five international benchmark tests covering social interaction, multi-subject behavior, three-dimensional movement analysis and autism behavior analysis. BehaVERT also features interpretability — it can tell researchers which specific behaviors it focused on when reaching a determination.

In experiments, the model focused on "nose-to-nose contact" when distinguishing autism model mice from normal mice. This aligns precisely with existing research showing that autism model mice perform approach behaviors normally but exhibit deficits in actual social interaction. The AI identified a core characteristic of autism-related behavior through observation alone, without having been trained on any prior biological knowledge.

The research team also confirmed that the AI understands the meaning of behaviors rather than merely classifying them. Inside the model, behavioral characteristics such as movement, attention and sociality were systematically organized — suggesting that animal behavior may contain a semantic structure similar to that of language.

"We can quantify how drug candidates change an animal's social, anxiety-related and motor behaviors," Professor Kim said. "It can be applied directly to behavioral phenotype analysis across a wide range of disease models, including autism, depression, schizophrenia and Parkinson's disease."

The findings were published March 24 in the International Journal of Computer Vision.