“In an Anthropic system card, internal activation analysis showed that the first token a model sees in a chat, "human:", registered with a negative valence.”