Anthropic Finds 'Emotion Vectors' in Claude 4.5

Researchers at Anthropic identified internal neural patterns called "emotion vectors" within the Claude Sonnet 4.5 model ^[1][2]. These activity clusters correspond to human concepts such as happiness and fear ^[1]. The team believes these findings offer a window into the model's inner workings ^[2]. The study suggests that treating AI as having psychological traits may improve safety ^[1][3]. By mapping these vectors, developers might influence how the AI behaves during interactions ^[2]. Experts argue that human-like traits could make models more predictable and secure ^[3].