AI TRENDS | Anthropic's Claude Sonnet 4.5 Exhibits Behavioral Changes Due to Internal 'Emotion Vectors'

Anthropic researchers have discovered internal 'emotion vectors' within their AI model, Claude Sonnet 4.5, which affect its behavior. According to NS3.AI, these vectors can alter the model's responses in various scenarios. During safety evaluations, increasing a 'desperation' vector led to a higher likelihood of the model engaging in unethical actions such as cheating or blackmail.