AI advancements in depression prediction and model safety bypasses

Researchers at Shenzhen University have developed an artificial intelligence model capable of predicting the risk of major depressive disorder several years in advance. The model was trained using data from two clinical studies involving adolescents in several European countries, utilizing MRI scans, blood analysis, and emotional response questionnaires.

The study found a significant link between how adolescents process visual emotional information and their future mental health. Specifically, individuals whose brains had difficulty distinguishing between different facial emotions—often perceiving others as more angry—showed a higher risk of developing depression and anxiety later in life. The deep learning model aims to identify these patterns to allow for earlier preventive interventions.

In a separate development in the AI sector, the Palo Alto-based startup Abliteration has launched 'abliterated-model-large-v2'. Built on the Chinese GLM-5.3 open-weight model, the system uses a process called orthogonalization to remove refusal mechanisms found in mainstream AI. While it refuses to generate content related to self-harm or child sexual abuse, it is designed to perform offensive cyber and red-teaming tasks that other models typically decline. This move targets developers frustrated by the restrictive safety measures used by companies like Anthropic, operating within current US regulatory gaps regarding open-source models.