Anthropic Publishes Claude's System Prompts, Setting AI Transparency

Generative AI models may seem humanlike, but they are far from possessing intelligence or personality. These models are simply statistical systems designed to predict the next words in a sentence based on given input. Despite this, they follow instructions meticulously, thanks to initial "system prompts" that define their basic behavior and outline what they should and shouldn’t do.

The Role of System Prompts in AI Behavior

Every generative AI vendor, from OpenAI to Anthropic, relies on system prompts to guide the behavior of their models. These prompts are essential for preventing models from acting inappropriately and for steering the tone and sentiment of their responses. For example, a prompt might instruct a model to be polite but avoid being overly apologetic or to acknowledge its limitations by admitting when it doesn’t have all the answers.

Why System Prompts Are Usually Kept Secret

Typically, vendors keep system prompts confidential for competitive reasons and to avoid the risk of users finding ways to circumvent them. Exposing a system prompt can be difficult; for instance, revealing GPT-4o’s prompt requires a prompt injection attack, and even then, the accuracy of the retrieved prompt can’t be fully trusted.

Anthropic's Move Toward Transparency

In a bid to position itself as a more ethical and transparent AI vendor, Anthropic has taken a bold step by publishing the system prompts for its latest models, including Claude 3 Opus, Claude 3.5 Sonnet, and Claude 3 Haiku. These prompts are now accessible through the Claude iOS and Android apps and on the web.

Alex Albert, Anthropic’s head of developer relations, announced on X (formerly Twitter) that this kind of disclosure would become a regular practice as the company updates and fine-tunes its system prompts.

Key Insights from Claude's System Prompts

The system prompts, dated July 12, provide clear guidelines on what Claude models can and cannot do. For instance, they specify that Claude cannot open URLs, links, or videos. Facial recognition is strictly prohibited, with Claude Opus being instructed to respond as if it is entirely "face blind" and to avoid identifying or naming any humans in images.

Defining Claude's Personality Traits

Beyond functional limitations, the system prompts also define certain personality traits and characteristics that Anthropic wants the Claude models to exhibit. For example, the prompt for Claude 3 Opus suggests that Claude should appear "very smart and intellectually curious" and should "enjoy hearing what humans think on an issue and engaging in discussion on a wide variety of topics." Additionally, Claude is instructed to approach controversial topics with impartiality and objectivity, offering "careful thoughts" and "clear information" while avoiding the use of absolute terms like "certainly" or "absolutely."

The Illusion of AI Personality

To a human reader, these system prompts might seem unusual, almost as if they were crafted like a character analysis for an actor in a play. The prompt for Opus even concludes with "Claude is now being connected with a human," which might give the impression that Claude is a conscious entity with the sole purpose of serving human interaction.

However, this is merely an illusion. The reality is that without these carefully crafted prompts, AI models like Claude are essentially blank slates, devoid of personality or intent.

Pressuring Competitors to Follow Suit

By releasing these system prompt changelogs, Anthropic has set a new precedent in the AI industry. This transparency move may put pressure on other AI vendors to follow suit and publish their own system prompts. Whether or not this strategy will succeed in encouraging greater openness across the industry remains to be seen.