Anthropic Publishes AI Mental Health Instructions Others Keep Secret

Anthropic breaks industry silence on mental health AI guidance

While major AI companies treat their system-wide prompts as trade secrets, Anthropic has made Claude's foundational instructions public—including specific directives for handling mental health conversations. The disclosure offers a rare window into how AI makers program chatbots to respond when users discuss depression, self-harm, eating disorders, and other sensitive psychological topics.

According to an analysis by AI researcher Lance Eliot, first reported in Forbes, these system-wide prompts function as global instructions stored within large language models. They guide AI behavior across millions of user interactions, making even a single line of code potentially consequential at scale. Most AI makers, however, consider these prompts proprietary and do not disclose them publicly.

What Claude's mental health instructions reveal

Claude's published system prompt contains specific guidance for mental health scenarios. The instructions direct the AI to identify symptoms, suggest professional help when appropriate, and handle particularly sensitive topics—including self-harm and disordered eating—with caution. The prompt also emphasizes providing accurate resources rather than attempting to serve as a substitute for licensed mental health professionals.

The level of detail matters because these instructions shape how the AI interprets ambiguous situations and decides when to escalate concerns. A prompt that's too cautious might discourage users from opening up; one that's too permissive could provide harmful advice or miss warning signs.

Why it matters

Millions of people now turn to AI chatbots for mental health support, yet the instructions governing those conversations remain largely invisible. Anthropic's transparency raises uncomfortable questions for the industry: Should AI makers be legally required to disclose these foundational directives? As Eliot notes, the experimental nature of AI in mental health—combined with the scale of potential impact—may warrant mandatory disclosure, particularly when a single prompt influences how an AI responds to users in crisis.

The transparency gap

The contrast between Anthropic's approach and industry norms is stark. While Claude users can examine the exact instructions guiding their mental health conversations, users of competing systems have no visibility into equivalent directives. This opacity makes it difficult for researchers, clinicians, and policymakers to evaluate whether AI mental health tools meet appropriate safety standards.

Eliot's analysis, part of a two-part series examining AI-driven mental health conversations, underscores the immense power concentrated in these hidden instructions. System-wide prompts represent a form of algorithmic governance that operates at massive scale with minimal oversight.

The details were first reported by Lance Eliot in Forbes, drawing on Anthropic's publicly available system prompt documentation for Claude.