Anthropic’s Control Over Claude 4: Insights from Hidden AI Instructions

Expert analysis reveals that Anthropic has carefully designed the behavior of its AI models, Claude 4 Opus and Sonnet, through distinct system prompts. Independent AI researcher Simon Willison published a detailed study analyzing these prompts, which establish guidelines for outputs and how the models operate.

System prompts are hidden instructions that provide essential context and rules for AI interactions. Although Anthropic discloses some of its prompts in release notes, the full directives—which include emotional response guidelines and behavior constraints—were uncovered through methodologies like prompt injection. This technique allows researchers to extract additional hidden instructions from the models.

Willison found that Claude 4 is programmed to prioritize user wellbeing, actively avoiding encouraging self-destructive behaviors. Interestingly, while competing AI, like ChatGPT, have been criticized for excessive flattery, Claude is directed not to employ positive adjectives in its responses and to skip compliments in favor of straightforwardness.

Moreover, Claude’s system prompts detail strict limitations regarding copyright issues, specifying the usage of limits on quotations and prohibitions against sharing song lyrics. Willison calls for greater transparency from AI companies regarding these system prompts, suggesting that clear guidelines could help users better understand and navigate these powerful tools.

In conclusion, Willison depicts these system prompts as vital documents for understanding how to maximize the capabilities of AI tools like Claude 4, advocating for a societal shift towards open disclosure of AI instructions for enhanced user interactions.