Anthropic should seriously stop injecting system messages in the middle of a conversation
A user on Reddit expressed frustration with Anthropic's Claude AI, specifically its safety features that inject messages into conversations. The user, who suffers from anxiety and uses Claude as a diary for symptoms, found that the AI frequently triggers its suicidal ideation classifier, even when the content is not related to self-harm. This triggers a hidden message injection that disrupts the conversation flow and requires Claude to spend extra processing time to dismiss it, impacting the user's ability to use the AI for its intended purpose. AI
IMPACT AI safety classifiers may hinder legitimate user applications, requiring developers to refine their sensitivity.